wasmtime 12.0.1 → 13.0.0

data/ext/cargo-vendor/cranelift-codegen-0.100.0/src/isa/riscv64/lower.isle

@@ -0,0 +1,2119 @@
+;; riscv64 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)
+
+;;;; Rules for `iconst` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty (iconst (u64_from_imm64 n))))
+  (imm ty n))
+
+;; ;;;; Rules for `vconst` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type (ty_vec_fits_in_register ty) (vconst n)))
+  (gen_constant ty (const_to_vconst n)))
+
+;;;; Rules for `f32const` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (f32const (u32_from_ieee32 n)))
+  (imm $F32 n))
+
+;;;; Rules for `f64const` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (f64const (u64_from_ieee64 n)))
+  (imm $F64 n))
+
+;;;; Rules for `null` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty (null)))
+  (imm ty 0))
+
+
+;;;; Rules for `iadd` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; Base case, simply adding things in registers.
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (iadd x y)))
+  (rv_add x y))
+
+;; Special cases for when one operand is an immediate that fits in 12 bits.
+(rule 1 (lower (has_type (ty_int_ref_scalar_64 ty) (iadd x (imm12_from_value y))))
+  (alu_rr_imm12 (select_addi ty) x y))
+
+(rule 2 (lower (has_type (ty_int_ref_scalar_64 ty) (iadd (imm12_from_value x) y)))
+  (alu_rr_imm12 (select_addi ty) y x))
+
+;; Special case when one of the operands is uextended
+;; Needs `Zba`
+(rule 3 (lower (has_type $I64 (iadd x (uextend y @ (value_type $I32)))))
+  (if-let $true (has_zba))
+  (rv_adduw y x))
+
+(rule 4 (lower (has_type $I64 (iadd (uextend x @ (value_type $I32)) y)))
+  (if-let $true (has_zba))
+  (rv_adduw x y))
+
+;; Add with const shift. We have a few of these instructions with `Zba`.
+(decl pure partial match_shnadd (Imm64) AluOPRRR)
+(rule (match_shnadd (u64_from_imm64 1)) (AluOPRRR.Sh1add))
+(rule (match_shnadd (u64_from_imm64 2)) (AluOPRRR.Sh2add))
+(rule (match_shnadd (u64_from_imm64 3)) (AluOPRRR.Sh3add))
+
+(rule 3 (lower (has_type $I64 (iadd x (ishl y (maybe_uextend (iconst n))))))
+  (if-let $true (has_zba))
+  (if-let shnadd (match_shnadd n))
+  (alu_rrr shnadd y x))
+
+(rule 4 (lower (has_type $I64 (iadd (ishl x (maybe_uextend (iconst n))) y)))
+  (if-let $true (has_zba))
+  (if-let shnadd (match_shnadd n))
+  (alu_rrr shnadd x y))
+
+
+;; Add with uextended const shift. We have a few of these instructions with `Zba`.
+;;
+;; !!! Important !!!
+;; These rules only work for (ishl (uextend _) _) and not for (uextend (ishl _ _))!
+;; Getting this wrong means a potential misscalculation of the shift amount.
+;; Additionaly we can only ensure that this is correct if the uextend is 32 to 64 bits.
+(decl pure partial match_shnadd_uw (Imm64) AluOPRRR)
+(rule (match_shnadd_uw (u64_from_imm64 1)) (AluOPRRR.Sh1adduw))
+(rule (match_shnadd_uw (u64_from_imm64 2)) (AluOPRRR.Sh2adduw))
+(rule (match_shnadd_uw (u64_from_imm64 3)) (AluOPRRR.Sh3adduw))
+
+(rule 5 (lower (has_type $I64 (iadd x (ishl (uextend y @ (value_type $I32)) (maybe_uextend (iconst n))))))
+  (if-let $true (has_zba))
+  (if-let shnadd_uw (match_shnadd_uw n))
+  (alu_rrr shnadd_uw y x))
+
+(rule 6 (lower (has_type $I64 (iadd (ishl (uextend x @ (value_type $I32)) (maybe_uextend (iconst n))) y)))
+  (if-let $true (has_zba))
+  (if-let shnadd_uw (match_shnadd_uw n))
+  (alu_rrr shnadd_uw x y))
+
+;; I128 cases
+(rule 7 (lower (has_type $I128 (iadd x y)))
+  (let ((low XReg (rv_add (value_regs_get x 0) (value_regs_get y 0)))
+        ;; compute carry.
+        (carry XReg (rv_sltu low (value_regs_get y 0)))
+        ;;
+        (high_tmp XReg (rv_add (value_regs_get x 1) (value_regs_get y 1)))
+        ;; add carry.
+        (high XReg (rv_add high_tmp carry)))
+    (value_regs low high)))
+
+;; SIMD Vectors
+(rule 8 (lower (has_type (ty_vec_fits_in_register ty) (iadd x y)))
+  (rv_vadd_vv x y (unmasked) ty))
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (splat y))))
+  (rv_vadd_vx x y (unmasked) ty))
+
+(rule 10 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (splat (sextend y @ (value_type sext_ty))))))
+  (if-let half_ty (ty_half_width ty))
+  (if-let $true (ty_equal (lane_type half_ty) sext_ty))
+  (rv_vwadd_wx x y (unmasked) (vstate_mf2 half_ty)))
+
+(rule 10 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (splat (uextend y @ (value_type uext_ty))))))
+  (if-let half_ty (ty_half_width ty))
+  (if-let $true (ty_equal (lane_type half_ty) uext_ty))
+  (rv_vwaddu_wx x y (unmasked) (vstate_mf2 half_ty)))
+
+(rule 11 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (replicated_imm5 y))))
+  (rv_vadd_vi x y (unmasked) ty))
+
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register ty) (iadd (splat x) y)))
+  (rv_vadd_vx y x (unmasked) ty))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register ty) (iadd (splat (sextend x @ (value_type sext_ty))) y)))
+  (if-let half_ty (ty_half_width ty))
+  (if-let $true (ty_equal (lane_type half_ty) sext_ty))
+  (rv_vwadd_wx y x (unmasked) (vstate_mf2 half_ty)))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register ty) (iadd (splat (uextend x @ (value_type uext_ty))) y)))
+  (if-let half_ty (ty_half_width ty))
+  (if-let $true (ty_equal (lane_type half_ty) uext_ty))
+  (rv_vwaddu_wx y x (unmasked) (vstate_mf2 half_ty)))
+
+(rule 14 (lower (has_type (ty_vec_fits_in_register ty) (iadd (replicated_imm5 x) y)))
+  (rv_vadd_vi y x (unmasked) ty))
+
+;; Signed Widening Low Additions
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register _) (iadd x (swiden_low y @ (value_type in_ty)))))
+  (rv_vwadd_wv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_low x @ (value_type in_ty)) y)))
+  (rv_vwadd_wv y x (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_low x @ (value_type in_ty))
+                                                            (swiden_low y))))
+  (rv_vwadd_vv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_low x @ (value_type in_ty))
+                                                            (splat (sextend y @ (value_type sext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) sext_ty))
+  (rv_vwadd_vx x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 15 (lower (has_type (ty_vec_fits_in_register _) (iadd (splat (sextend x @ (value_type sext_ty)))
+                                                            (swiden_low y @ (value_type in_ty)))))
+  (if-let $true (ty_equal (lane_type in_ty) sext_ty))
+  (rv_vwadd_vx y x (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Signed Widening High Additions
+;; These are the same as the low additions, but we first slide down the inputs.
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register _) (iadd x (swiden_high y @ (value_type in_ty)))))
+  (rv_vwadd_wv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_high x @ (value_type in_ty)) y)))
+  (rv_vwadd_wv y (gen_slidedown_half in_ty x) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_high x @ (value_type in_ty))
+                                                            (swiden_high y))))
+  (rv_vwadd_vv (gen_slidedown_half in_ty x) (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_high x @ (value_type in_ty))
+                                                            (splat (sextend y @ (value_type sext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) sext_ty))
+  (rv_vwadd_vx (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 15 (lower (has_type (ty_vec_fits_in_register _) (iadd (splat (sextend x @ (value_type sext_ty)))
+                                                            (swiden_high y @ (value_type in_ty)))))
+  (if-let $true (ty_equal (lane_type in_ty) sext_ty))
+  (rv_vwadd_vx (gen_slidedown_half in_ty y) x (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Unsigned Widening Low Additions
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register _) (iadd x (uwiden_low y @ (value_type in_ty)))))
+  (rv_vwaddu_wv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_low x @ (value_type in_ty)) y)))
+  (rv_vwaddu_wv y x (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_low x @ (value_type in_ty))
+                                                            (uwiden_low y))))
+  (rv_vwaddu_vv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_low x @ (value_type in_ty))
+                                                            (splat (uextend y @ (value_type uext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) uext_ty))
+  (rv_vwaddu_vx x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 15 (lower (has_type (ty_vec_fits_in_register _) (iadd (splat (uextend x @ (value_type uext_ty)))
+                                                            (uwiden_low y @ (value_type in_ty)))))
+  (if-let $true (ty_equal (lane_type in_ty) uext_ty))
+  (rv_vwaddu_vx y x (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Unsigned Widening High Additions
+;; These are the same as the low additions, but we first slide down the inputs.
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register _) (iadd x (uwiden_high y @ (value_type in_ty)))))
+  (rv_vwaddu_wv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_high x @ (value_type in_ty)) y)))
+  (rv_vwaddu_wv y (gen_slidedown_half in_ty x) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_high x @ (value_type in_ty))
+                                                            (uwiden_high y))))
+  (rv_vwaddu_vv (gen_slidedown_half in_ty x) (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_high x @ (value_type in_ty))
+                                                            (splat (uextend y @ (value_type uext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) uext_ty))
+  (rv_vwaddu_vx (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 15 (lower (has_type (ty_vec_fits_in_register _) (iadd (splat (uextend y @ (value_type uext_ty)))
+                                                            (uwiden_high x @ (value_type in_ty)))))
+  (if-let $true (ty_equal (lane_type in_ty) uext_ty))
+  (rv_vwaddu_vx (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Signed Widening Mixed High/Low Additions
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_low x @ (value_type in_ty))
+                                                            (swiden_high y))))
+  (rv_vwadd_vv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (swiden_high x @ (value_type in_ty))
+                                                            (swiden_low y))))
+  (rv_vwadd_vv (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Unsigned Widening Mixed High/Low Additions
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_low x @ (value_type in_ty))
+                                                            (uwiden_high y))))
+  (rv_vwaddu_vv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register _) (iadd (uwiden_high x @ (value_type in_ty))
+                                                            (uwiden_low y))))
+  (rv_vwaddu_vv (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Fused Multiply Accumulate Rules `vmacc`
+;;
+;; I dont think we can use `vmadd`/`vmnsub` here since it just modifies the multiplication
+;; register instead of the addition one. The actual pattern matched seems to be
+;; exactly the same.
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (imul y z))))
+  (rv_vmacc_vv x y z (unmasked) ty))
+
+(rule 10 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (imul y (splat z)))))
+  (rv_vmacc_vx x y z (unmasked) ty))
+
+(rule 11 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (imul (splat y) z))))
+  (rv_vmacc_vx x z y (unmasked) ty))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register ty) (iadd (imul x y) z)))
+  (rv_vmacc_vv z x y (unmasked) ty))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register ty) (iadd (imul x (splat y)) z)))
+  (rv_vmacc_vx z x y (unmasked) ty))
+
+(rule 14 (lower (has_type (ty_vec_fits_in_register ty) (iadd (imul (splat x) y) z)))
+  (rv_vmacc_vx z y x (unmasked) ty))
+
+;; Fused Multiply Subtract Rules `vnmsac`
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (ineg (imul y z)))))
+  (rv_vnmsac_vv x y z (unmasked) ty))
+
+(rule 10 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (ineg (imul y (splat z))))))
+  (rv_vnmsac_vx x y z (unmasked) ty))
+
+(rule 11 (lower (has_type (ty_vec_fits_in_register ty) (iadd x (ineg (imul (splat y) z)))))
+  (rv_vnmsac_vx x z y (unmasked) ty))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register ty) (iadd (ineg (imul x y)) z)))
+  (rv_vnmsac_vv z x y (unmasked) ty))
+
+(rule 13 (lower (has_type (ty_vec_fits_in_register ty) (iadd (ineg (imul x (splat y))) z)))
+  (rv_vnmsac_vx z x y (unmasked) ty))
+
+(rule 14 (lower (has_type (ty_vec_fits_in_register ty) (iadd (ineg (imul (splat x) y)) z)))
+  (rv_vnmsac_vx z y x (unmasked) ty))
+
+;;; Rules for `uadd_overflow_trap` ;;;;;;;;;;;;;
+(rule
+  (lower (has_type (fits_in_64 ty) (uadd_overflow_trap x y tc)))
+  (let ((res ValueRegs (lower_uadd_overflow x y ty))
+        (_ InstOutput (gen_trapif (value_regs_get res 1) tc)))
+    (value_regs_get res 0)))
+
+
+;;;; Rules for `isub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Base case, simply subtracting things in registers.
+
+(rule (lower (has_type (ty_int_ref_scalar_64 ty) (isub x y)))
+  (rv_sub x y))
+
+(rule 1 (lower (has_type (fits_in_32 (ty_int ty)) (isub x y)))
+  (rv_subw x y))
+
+(rule 2 (lower (has_type $I128 (isub x y)))
+  (i128_sub x y))
+
+;; SIMD Vectors
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (isub x y)))
+  (rv_vsub_vv x y (unmasked) ty))
+
+(rule 4 (lower (has_type (ty_vec_fits_in_register ty) (isub x (splat y))))
+  (rv_vsub_vx x y (unmasked) ty))
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (isub x (splat (sextend y @ (value_type sext_ty))))))
+  (if-let half_ty (ty_half_width ty))
+  (if-let $true (ty_equal (lane_type half_ty) sext_ty))
+  (rv_vwsub_wx x y (unmasked) (vstate_mf2 half_ty)))
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (isub x (splat (uextend y @ (value_type uext_ty))))))
+  (if-let half_ty (ty_half_width ty))
+  (if-let $true (ty_equal (lane_type half_ty) uext_ty))
+  (rv_vwsubu_wx x y (unmasked) (vstate_mf2 half_ty)))
+
+(rule 6 (lower (has_type (ty_vec_fits_in_register ty) (isub (splat x) y)))
+  (rv_vrsub_vx y x (unmasked) ty))
+
+(rule 7 (lower (has_type (ty_vec_fits_in_register ty) (isub (replicated_imm5 x) y)))
+  (rv_vrsub_vi y x (unmasked) ty))
+
+
+;; Signed Widening Low Subtractions
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register _) (isub x (swiden_low y @ (value_type in_ty)))))
+  (rv_vwsub_wv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (swiden_low x @ (value_type in_ty))
+                                                           (swiden_low y))))
+  (rv_vwsub_vv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (swiden_low x @ (value_type in_ty))
+                                                           (splat (sextend y @ (value_type sext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) sext_ty))
+  (rv_vwsub_vx x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Signed Widening High Subtractions
+;; These are the same as the low widenings, but we first slide down the inputs.
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register _) (isub x (swiden_high y @ (value_type in_ty)))))
+  (rv_vwsub_wv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (swiden_high x @ (value_type in_ty))
+                                                           (swiden_high y))))
+  (rv_vwsub_vv (gen_slidedown_half in_ty x) (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (swiden_high x @ (value_type in_ty))
+                                                           (splat (sextend y @ (value_type sext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) sext_ty))
+  (rv_vwsub_vx (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Unsigned Widening Low Subtractions
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register _) (isub x (uwiden_low y @ (value_type in_ty)))))
+  (rv_vwsubu_wv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (uwiden_low x @ (value_type in_ty))
+                                                           (uwiden_low y))))
+  (rv_vwsubu_vv x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (uwiden_low x @ (value_type in_ty))
+                                                           (splat (uextend y @ (value_type uext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) uext_ty))
+  (rv_vwsubu_vx x y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Unsigned Widening High Subtractions
+;; These are the same as the low widenings, but we first slide down the inputs.
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register _) (isub x (uwiden_high y @ (value_type in_ty)))))
+  (rv_vwsubu_wv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (uwiden_high x @ (value_type in_ty))
+                                                           (uwiden_high y))))
+  (rv_vwsubu_vv (gen_slidedown_half in_ty x) (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (uwiden_high x @ (value_type in_ty))
+                                                           (splat (uextend y @ (value_type uext_ty))))))
+  (if-let $true (ty_equal (lane_type in_ty) uext_ty))
+  (rv_vwsubu_vx (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Signed Widening Mixed High/Low Subtractions
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (swiden_low x @ (value_type in_ty))
+                                                           (swiden_high y))))
+  (rv_vwsub_vv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (swiden_high x @ (value_type in_ty))
+                                                           (swiden_low y))))
+  (rv_vwsub_vv (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+;; Unsigned Widening Mixed High/Low Subtractions
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (uwiden_low x @ (value_type in_ty))
+                                                           (uwiden_high y))))
+  (rv_vwsubu_vv x (gen_slidedown_half in_ty y) (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register _) (isub (uwiden_high x @ (value_type in_ty))
+                                                           (uwiden_low y))))
+  (rv_vwsubu_vv (gen_slidedown_half in_ty x) y (unmasked) (vstate_mf2 (ty_half_lanes in_ty))))
+
+
+;;;; Rules for `ineg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type (ty_int ty) (ineg val)))
+  (neg ty val))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (ineg x)))
+  (rv_vneg_v x (unmasked) ty))
+
+
+;;;; Rules for `imul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (imul x y)))
+  (rv_mul x y))
+
+(rule 1 (lower (has_type (fits_in_32 (ty_int ty)) (imul x y)))
+  (rv_mulw x y))
+
+;; for I128
+(rule 2 (lower (has_type $I128 (imul x y)))
+  (let
+    ((x_regs ValueRegs x)
+      (x_lo XReg (value_regs_get x_regs 0))
+      (x_hi XReg (value_regs_get x_regs 1))
+
+      ;; Get the high/low registers for `y`.
+      (y_regs ValueRegs y)
+      (y_lo XReg (value_regs_get y_regs 0))
+      (y_hi XReg (value_regs_get y_regs 1))
+
+      ;; 128bit mul formula:
+      ;;   dst_lo = x_lo * y_lo
+      ;;   dst_hi = mulhu(x_lo, y_lo) + (x_lo * y_hi) + (x_hi * y_lo)
+      ;;
+      ;; We can convert the above formula into the following
+      ;; mulhu   dst_hi, x_lo, y_lo
+      ;; madd    dst_hi, x_lo, y_hi, dst_hi
+      ;; madd    dst_hi, x_hi, y_lo, dst_hi
+      ;; madd    dst_lo, x_lo, y_lo, zero
+      (dst_hi1 XReg (rv_mulhu x_lo y_lo))
+      (dst_hi2 XReg (madd x_lo y_hi dst_hi1))
+      (dst_hi XReg (madd x_hi y_lo dst_hi2))
+      (dst_lo XReg (madd x_lo y_lo (zero_reg))))
+    (value_regs dst_lo dst_hi)))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (imul x y)))
+  (rv_vmul_vv x y (unmasked) ty))
+
+(rule 4 (lower (has_type (ty_vec_fits_in_register ty) (imul (splat x) y)))
+  (rv_vmul_vx y x (unmasked) ty))
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (imul x (splat y))))
+  (rv_vmul_vx x y (unmasked) ty))
+
+;;;; Rules for `smulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (smulhi x y)))
+  (lower_smlhi ty (sext x ty $I64) (sext y ty $I64)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (smulhi x y)))
+  (rv_vmulh_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (smulhi (splat x) y)))
+  (rv_vmulh_vx y x (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (smulhi x (splat y))))
+  (rv_vmulh_vx x y (unmasked) ty))
+
+;;;; Rules for `umulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (umulhi x y)))
+  (lower_umlhi ty (zext x ty $I64) (zext y ty $I64)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (umulhi x y)))
+  (rv_vmulhu_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (umulhi (splat x) y)))
+  (rv_vmulhu_vx y x (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (umulhi x (splat y))))
+  (rv_vmulhu_vx x y (unmasked) ty))
+
+;;;; Rules for `div` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule -1 (lower (has_type (fits_in_32 ty) (udiv x y)))
+  (let
+    ((y2 XReg (zext y ty $I64))
+      (_ InstOutput (gen_div_by_zero y2)))
+    (rv_divuw (zext x ty $I64) y2)))
+
+(rule -1 (lower (has_type (fits_in_32 ty) (sdiv x y)))
+  (let
+    ((a XReg (sext x ty $I64))
+      (b XReg (sext y ty $I64))
+      (_ InstOutput (gen_div_overflow a b ty))
+      (_ InstOutput (gen_div_by_zero b)))
+    (rv_divw a b)))
+
+(rule (lower (has_type $I64 (sdiv x y)))
+  (let
+    ((_ InstOutput (gen_div_overflow x y $I64))
+      (_ InstOutput (gen_div_by_zero y))    )
+    (rv_div x y)))
+
+(rule (lower (has_type $I64 (udiv x y)))
+  (let
+    ((_ InstOutput (gen_div_by_zero y)))
+    (rv_divu x y)))
+
+;;;; Rules for `rem` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule -1 (lower (has_type (fits_in_16 ty) (urem x y)))
+  (let
+    ((y2 XReg (zext y ty $I64))
+      (_ InstOutput (gen_div_by_zero y2)))
+    (rv_remuw (zext x ty $I64) y2)))
+
+(rule -1 (lower (has_type (fits_in_16 ty) (srem x y)))
+  (let
+    ((y2 XReg (sext y ty $I64))
+      (_ InstOutput (gen_div_by_zero y2)))
+    (rv_remw (sext x ty $I64) y2)))
+
+(rule (lower (has_type $I32 (srem x y)))
+  (let
+    ((y2 XReg (sext y $I32 $I64))
+      (_ InstOutput (gen_div_by_zero y2)))
+   (rv_remw x y2)))
+
+(rule (lower (has_type $I32 (urem x y)))
+  (let
+    ((y2 XReg (zext y $I32 $I64))
+        (_ InstOutput (gen_div_by_zero y2)))
+    (rv_remuw x y2)))
+
+(rule (lower (has_type $I64 (srem x y)))
+  (let
+    ((_ InstOutput (gen_div_by_zero y)))
+    (rv_rem x y)))
+
+(rule (lower (has_type $I64 (urem x y)))
+  (let
+    ((_ InstOutput (gen_div_by_zero y)))
+    (rv_remu x y)))
+
+;;;; Rules for `and` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_int ty) (band x y)))
+  (gen_and ty x y))
+
+;; Special cases for when one operand is an immediate that fits in 12 bits.
+(rule 1 (lower (has_type (fits_in_64 (ty_int ty)) (band x (imm12_from_value y))))
+  (rv_andi x y))
+
+(rule 2 (lower (has_type (fits_in_64 (ty_int ty)) (band (imm12_from_value x) y)))
+  (rv_andi y x))
+
+(rule 3 (lower (has_type (ty_scalar_float ty) (band x y)))
+  (lower_float_binary (AluOPRRR.And) x y ty))
+
+;; 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.
+
+(rule 4 (lower (has_type (fits_in_64 (ty_int ty)) (band x (bnot y))))
+  (if-let $true (has_zbb))
+  (rv_andn x y))
+
+(rule 5 (lower (has_type (fits_in_64 (ty_int ty)) (band (bnot y) x)))
+  (if-let $true (has_zbb))
+  (rv_andn x y))
+
+(rule 6 (lower (has_type $I128 (band x (bnot y))))
+  (if-let $true (has_zbb))
+  (let ((low XReg (rv_andn (value_regs_get x 0) (value_regs_get y 0)))
+        (high XReg (rv_andn (value_regs_get x 1) (value_regs_get y 1))))
+    (value_regs low high)))
+
+(rule 7 (lower (has_type $I128 (band (bnot y) x)))
+  (if-let $true (has_zbb))
+  (let ((low XReg (rv_andn (value_regs_get x 0) (value_regs_get y 0)))
+        (high XReg (rv_andn (value_regs_get x 1) (value_regs_get y 1))))
+    (value_regs low high)))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register ty) (band x y)))
+  (rv_vand_vv x y (unmasked) ty))
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register ty) (band x (splat y))))
+  (if (ty_vector_not_float ty))
+  (rv_vand_vx x y (unmasked) ty))
+
+(rule 10 (lower (has_type (ty_vec_fits_in_register ty) (band (splat x) y)))
+  (if (ty_vector_not_float ty))
+  (rv_vand_vx y x (unmasked) ty))
+
+(rule 11 (lower (has_type (ty_vec_fits_in_register ty) (band x (replicated_imm5 y))))
+  (rv_vand_vi x y (unmasked) ty))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register ty) (band (replicated_imm5 x) y)))
+  (rv_vand_vi y x (unmasked) ty))
+
+
+;;;; Rules for `or` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_int ty) (bor x y)))
+  (gen_or ty x y))
+
+;; Special cases for when one operand is an immediate that fits in 12 bits.
+(rule 1 (lower (has_type (fits_in_64 (ty_int ty)) (bor x (imm12_from_value y))))
+  (rv_ori x y))
+
+(rule 2 (lower (has_type (fits_in_64 (ty_int ty)) (bor (imm12_from_value x) y)))
+  (rv_ori y x))
+
+(rule 3 (lower (has_type (ty_scalar_float ty) (bor x y)))
+  (lower_float_binary (AluOPRRR.Or) x y ty))
+
+;; Specialized lowerings for `(bor x (bnot y))` which is additionally produced
+;; by Cranelift's `bor_not` instruction that is legalized into the simpler
+;; forms early on.
+
+(rule 4 (lower (has_type (fits_in_64 (ty_int ty)) (bor x (bnot y))))
+  (if-let $true (has_zbb))
+  (rv_orn x y))
+
+(rule 5 (lower (has_type (fits_in_64 (ty_int ty)) (bor (bnot y) x)))
+  (if-let $true (has_zbb))
+  (rv_orn x y))
+
+(rule 6 (lower (has_type $I128 (bor x (bnot y))))
+  (if-let $true (has_zbb))
+  (let ((low XReg (rv_orn (value_regs_get x 0) (value_regs_get y 0)))
+        (high XReg (rv_orn (value_regs_get x 1) (value_regs_get y 1))))
+    (value_regs low high)))
+
+(rule 7 (lower (has_type $I128 (bor (bnot y) x)))
+  (if-let $true (has_zbb))
+  (let ((low XReg (rv_orn (value_regs_get x 0) (value_regs_get y 0)))
+        (high XReg (rv_orn (value_regs_get x 1) (value_regs_get y 1))))
+    (value_regs low high)))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register ty) (bor x y)))
+  (rv_vor_vv x y (unmasked) ty))
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register ty) (bor x (splat y))))
+  (if (ty_vector_not_float ty))
+  (rv_vor_vx x y (unmasked) ty))
+
+(rule 10 (lower (has_type (ty_vec_fits_in_register ty) (bor (splat x) y)))
+  (if (ty_vector_not_float ty))
+  (rv_vor_vx y x (unmasked) ty))
+
+(rule 11 (lower (has_type (ty_vec_fits_in_register ty) (bor x (replicated_imm5 y))))
+  (rv_vor_vi x y (unmasked) ty))
+
+(rule 12 (lower (has_type (ty_vec_fits_in_register ty) (bor (replicated_imm5 x) y)))
+  (rv_vor_vi y x (unmasked) ty))
+
+
+;;;; Rules for `xor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (fits_in_64 (ty_int ty)) (bxor x y)))
+  (rv_xor x y))
+
+;; Special cases for when one operand is an immediate that fits in 12 bits.
+(rule 1 (lower (has_type (fits_in_64 (ty_int ty)) (bxor x (imm12_from_value y))))
+  (rv_xori x y))
+
+(rule 2 (lower (has_type (fits_in_64 (ty_int ty)) (bxor (imm12_from_value x) y)))
+  (rv_xori y x))
+
+(rule 3 (lower (has_type $I128 (bxor x y)))
+  (lower_b128_binary (AluOPRRR.Xor) x y))
+
+(rule 4 (lower (has_type (ty_scalar_float ty) (bxor x y)))
+  (lower_float_binary (AluOPRRR.Xor) x y ty))
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (bxor x y)))
+  (rv_vxor_vv x y (unmasked) ty))
+
+(rule 6 (lower (has_type (ty_vec_fits_in_register ty) (bxor x (splat y))))
+  (if (ty_vector_not_float ty))
+  (rv_vxor_vx x y (unmasked) ty))
+
+(rule 7 (lower (has_type (ty_vec_fits_in_register ty) (bxor (splat x) y)))
+  (if (ty_vector_not_float ty))
+  (rv_vxor_vx y x (unmasked) ty))
+
+(rule 8 (lower (has_type (ty_vec_fits_in_register ty) (bxor x (replicated_imm5 y))))
+  (rv_vxor_vi x y (unmasked) ty))
+
+(rule 9 (lower (has_type (ty_vec_fits_in_register ty) (bxor (replicated_imm5 x) y)))
+  (rv_vxor_vi y x (unmasked) ty))
+
+
+;;;; Rules for `bnot` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar ty) (bnot x)))
+  (gen_bnot ty x))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (bnot x)))
+  (rv_vnot_v x (unmasked) ty))
+
+;;;; Rules for `bit_reverse` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type (fits_in_64 (ty_int ty)) (bitrev x)))
+  (lower_bit_reverse x ty))
+
+(rule 1 (lower (has_type $I128 (bitrev x)))
+  (let ((val ValueRegs x)
+    (lo_rev XReg (lower_bit_reverse (value_regs_get val 0) $I64))
+    (hi_rev XReg (lower_bit_reverse (value_regs_get val 1) $I64)))
+    (value_regs hi_rev lo_rev)))
+
+;;;; Rules for `bswap` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 1 (lower (has_type (fits_in_64 (ty_int ty)) (bswap x)))
+  (gen_bswap ty x))
+
+(rule 2 (lower (has_type $I128 (bswap x)))
+  (value_regs
+    (gen_bswap $I64 (value_regs_get x 1))
+    (gen_bswap $I64 (value_regs_get x 0))))
+
+
+;;;; Rules for `ctz` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type (fits_in_64 ty) (ctz x)))
+  (lower_ctz ty x))
+
+(rule 1 (lower (has_type $I128 (ctz x)))
+  (lower_ctz_128 x))
+
+;;;; Rules for `clz` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type (fits_in_64 ty) (clz x)))
+  (lower_clz ty x))
+
+(rule 1 (lower (has_type $I128 (clz x)))
+  (lower_clz_i128 x))
+
+;;;; Rules for `cls` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type (fits_in_64 ty) (cls x)))
+  (lower_cls ty x))
+
+(rule 1 (lower (has_type $I128 (cls x)))
+  (lower_cls_i128 x))
+
+;;;; Rules for `uextend` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type out_ty (uextend val @ (value_type in_ty))))
+  (extend val (ExtendOp.Zero) in_ty out_ty))
+
+;;;; Rules for `sextend` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type out_ty (sextend val @ (value_type in_ty))))
+  (extend val (ExtendOp.Signed) in_ty out_ty))
+
+;; The instructions below are present in RV64I and sign-extend the result to 64 bits.
+
+(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (iadd x y)))))
+  (rv_addw x y))
+
+(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (isub x y)))))
+  (rv_subw x y))
+
+(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (ishl x y)))))
+  (rv_sllw x (value_regs_get y 0)))
+
+(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (ushr x y)))))
+  (rv_srlw x (value_regs_get y 0)))
+
+(rule 1 (lower (has_type $I64 (sextend (has_type $I32 (sshr x y)))))
+  (rv_sraw x (value_regs_get y 0)))
+
+
+(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (iadd x (imm12_from_value y))))))
+  (rv_addiw x y))
+
+(rule 3 (lower (has_type $I64 (sextend (has_type $I32 (iadd (imm12_from_value x) y)))))
+  (rv_addiw y x))
+
+(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (ishl x (imm12_from_value y))))))
+  (rv_slliw x y))
+
+(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (ushr x (imm12_from_value y))))))
+  (rv_srliw x y))
+
+(rule 2 (lower (has_type $I64 (sextend (has_type $I32 (sshr x (imm12_from_value y))))))
+  (rv_sraiw x y))
+
+;;;; Rules for `popcnt` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (popcnt x)))
+  (lower_popcnt x ty))
+
+(rule 1 (lower (has_type $I128 (popcnt x)))
+  (lower_popcnt_i128 x))
+
+;; Popcount using multiply.
+;; This is popcount64c() from
+;; http://en.wikipedia.org/wiki/Hamming_weight
+;;
+;; Here's the C version for 32 bits:
+;;  x = x - ((x>> 1) & 0x55555555);
+;;  x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
+;;  x = ((x + (x >> 4)) & 0x0F0F0F0F);
+;;  return (x * 0x01010101) >> 24; // Here 24 is the type width - 8.
+;;
+;; TODO: LLVM generates a much better implementation for I8X16. See: https://godbolt.org/z/qr6vf9Gr3
+;; For the other types it seems to be largely the same.
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (popcnt x)))
+  (if-let one (u64_to_uimm5 1))
+  (if-let two (u64_to_uimm5 2))
+  (if-let four (u64_to_uimm5 4))
+
+  (let (;; x = x - ((x >> 1) & 0x55555555);
+        (mask_55 XReg (imm (lane_type ty) (u64_and 0x5555555555555555 (ty_mask (lane_type ty)))))
+        (count2_shr VReg (rv_vsrl_vi x one (unmasked) ty))
+        (count2_and VReg (rv_vand_vx count2_shr mask_55 (unmasked) ty))
+        (count2 VReg (rv_vsub_vv x count2_and (unmasked) ty))
+
+        ;; x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
+        (mask_33 XReg (imm (lane_type ty) (u64_and 0x3333333333333333 (ty_mask (lane_type ty)))))
+        (count4_shr VReg (rv_vsrl_vi count2 two (unmasked) ty))
+        (count4_and VReg (rv_vand_vx count4_shr mask_33 (unmasked) ty))
+        (count4_lhs VReg (rv_vand_vx count2 mask_33 (unmasked) ty))
+        (count4 VReg (rv_vadd_vv count4_lhs count4_and (unmasked) ty))
+
+        ;; x = (x + (x >> 4)) & 0x0F0F0F0F;
+        (mask_0f XReg (imm (lane_type ty) (u64_and 0x0f0f0f0f0f0f0f0f (ty_mask (lane_type ty)))))
+        (count8_shr VReg (rv_vsrl_vi count4 four (unmasked) ty))
+        (count8_add VReg (rv_vadd_vv count4 count8_shr (unmasked) ty))
+        (count8 VReg (rv_vand_vx count8_add mask_0f (unmasked) ty))
+
+        ;; (x * 0x01010101) >> (<ty_width> - 8)
+        (mask_01 XReg (imm (lane_type ty) (u64_and 0x0101010101010101 (ty_mask (lane_type ty)))))
+        (mul VReg (rv_vmul_vx count8 mask_01 (unmasked) ty))
+        (shift XReg (imm $I64 (u64_sub (ty_bits (lane_type ty)) 8)))
+        (res VReg (rv_vsrl_vx mul shift (unmasked) ty)))
+    res))
+
+;;;; Rules for `ishl` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; 8/16 bit types need a mask on the shift amount
+(rule 0 (lower (has_type (ty_int (ty_8_or_16 ty)) (ishl x y)))
+  (if-let mask (u64_to_imm12 (shift_mask ty)))
+  (rv_sllw x (rv_andi (value_regs_get y 0) mask)))
+
+;; Using the 32bit version of `sll` automatically masks the shift amount.
+(rule 1 (lower (has_type $I32 (ishl x y)))
+  (rv_sllw x (value_regs_get y 0)))
+
+;; Similarly, the 64bit version does the right thing.
+(rule 1 (lower (has_type $I64 (ishl x y)))
+  (rv_sll x (value_regs_get y 0)))
+
+;; If the shift amount is known. We can mask it and encode it in the instruction.
+(rule 2 (lower (has_type (int_fits_in_32 ty) (ishl x (maybe_uextend (imm12_from_value y)))))
+  (rv_slliw x (imm12_and y (shift_mask ty))))
+
+;; We technically don't need to mask the shift amount here. The instruction
+;; does the right thing. But it's neater when pretty printing it.
+(rule 3 (lower (has_type ty @ $I64 (ishl x (maybe_uextend (imm12_from_value y)))))
+  (rv_slli x (imm12_and y (shift_mask ty))))
+
+;; With `Zba` we have a shift that zero extends the LHS argument.
+(rule 4 (lower (has_type $I64 (ishl (uextend x @ (value_type $I32)) (maybe_uextend (imm12_from_value y)))))
+  (if-let $true (has_zba))
+  (rv_slliuw x y))
+
+;; I128 cases
+(rule 4 (lower (has_type $I128 (ishl x y)))
+  (let ((tmp ValueRegs (gen_shamt $I128 (value_regs_get y 0)))
+        (shamt XReg (value_regs_get tmp 0))
+        (len_sub_shamt XReg (value_regs_get tmp 1))
+        ;;
+        (low XReg (rv_sll (value_regs_get x 0) shamt))
+        ;; high part.
+        (high_part1 XReg (rv_srl (value_regs_get x 0) len_sub_shamt))
+        (high_part2 XReg (gen_select_reg (IntCC.Equal) shamt (zero_reg) (zero_reg) high_part1))
+        ;;
+        (high_part3 XReg (rv_sll (value_regs_get x 1) shamt))
+        (high XReg (rv_or high_part2 high_part3))
+        ;;
+        (const64 XReg (imm $I64 64))
+        (shamt_128 XReg (rv_andi (value_regs_get y 0) (imm12_const 127))))
+    (value_regs
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 (zero_reg) low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 low high))))
+
+;; SIMD Cases
+;; We don't need to mask anything since it is done by the instruction according to SEW.
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (ishl x y)))
+  (rv_vsll_vx x (value_regs_get y 0) (unmasked) ty))
+
+(rule 6 (lower (has_type (ty_vec_fits_in_register ty) (ishl x (maybe_uextend (uimm5_from_value y)))))
+  (rv_vsll_vi x y (unmasked) ty))
+
+;;;; Rules for `ushr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; 8/16 bit types need a mask on the shift amount, and the LHS needs to be
+;; zero extended.
+(rule 0 (lower (has_type (ty_int (fits_in_16 ty)) (ushr x y)))
+  (if-let mask (u64_to_imm12 (shift_mask ty)))
+  (rv_srlw (zext x ty $I64) (rv_andi (value_regs_get y 0) mask)))
+
+;; Using the 32bit version of `srl` automatically masks the shift amount.
+(rule 1 (lower (has_type $I32 (ushr x y)))
+  (rv_srlw x (value_regs_get y 0)))
+
+;; Similarly, the 64bit version does the right thing.
+(rule 1 (lower (has_type $I64 (ushr x y)))
+  (rv_srl x (value_regs_get y 0)))
+
+;; When the RHS is known we can just encode it in the instruction.
+(rule 2 (lower (has_type (ty_int (fits_in_16 ty)) (ushr x (maybe_uextend (imm12_from_value y)))))
+  (rv_srliw (zext x ty $I64) (imm12_and y (shift_mask ty))))
+
+(rule 3 (lower (has_type $I32 (ushr x (maybe_uextend (imm12_from_value y)))))
+  (rv_srliw x y))
+
+(rule 3 (lower (has_type $I64 (ushr x (maybe_uextend (imm12_from_value y)))))
+  (rv_srli x y))
+
+(rule 3 (lower (has_type $I128 (ushr x y)))
+  (let ((tmp ValueRegs (gen_shamt $I128 (value_regs_get y 0)))
+        (shamt XReg (value_regs_get tmp 0))
+        (len_sub_shamt XReg (value_regs_get tmp 1))
+        ;; low part.
+        (low_part1 XReg (rv_sll (value_regs_get x 1) len_sub_shamt))
+        (low_part2 XReg (gen_select_reg (IntCC.Equal) shamt (zero_reg) (zero_reg) low_part1))
+        ;;
+        (low_part3 XReg (rv_srl (value_regs_get x 0) shamt))
+        (low XReg (rv_or low_part2 low_part3))
+        ;;
+        (const64 XReg (imm $I64 64))
+        ;;
+        (high XReg (rv_srl (value_regs_get x 1) shamt))
+        (shamt_128 XReg (rv_andi (value_regs_get y 0) (imm12_const 127))))
+    (value_regs
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 (zero_reg) high))))
+
+;; SIMD Cases
+;; We don't need to mask or extend anything since it is done by the instruction according to SEW.
+
+(rule 4 (lower (has_type (ty_vec_fits_in_register ty) (ushr x y)))
+  (rv_vsrl_vx x (value_regs_get y 0) (unmasked) ty))
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (ushr x (maybe_uextend (uimm5_from_value y)))))
+  (rv_vsrl_vi x y (unmasked) ty))
+
+;;;; Rules for `sshr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; 8/16 bit types need a mask on the shift amount, and the LHS needs to be
+;; zero extended.
+(rule 0 (lower (has_type (ty_int (fits_in_16 ty)) (sshr x y)))
+  (if-let mask (u64_to_imm12 (shift_mask ty)))
+  (rv_sraw (sext x ty $I64) (rv_andi (value_regs_get y 0) mask)))
+
+;; Using the 32bit version of `sra` automatically masks the shift amount.
+(rule 1 (lower (has_type $I32 (sshr x y)))
+  (rv_sraw x (value_regs_get y 0)))
+
+;; Similarly, the 64bit version does the right thing.
+(rule 1 (lower (has_type $I64 (sshr x y)))
+  (rv_sra x (value_regs_get y 0)))
+
+;; When the RHS is known we can just encode it in the instruction.
+(rule 2 (lower (has_type (ty_int (fits_in_16 ty)) (sshr x (maybe_uextend (imm12_from_value y)))))
+  (rv_sraiw (sext x ty $I64) (imm12_and y (shift_mask ty))))
+
+(rule 3 (lower (has_type $I32 (sshr x (maybe_uextend (imm12_from_value y)))))
+  (rv_sraiw x y))
+
+(rule 3 (lower (has_type $I64 (sshr x (maybe_uextend (imm12_from_value y)))))
+  (rv_srai x y))
+
+(rule 3 (lower (has_type $I128 (sshr x y)))
+  (let ((tmp ValueRegs (gen_shamt $I128 (value_regs_get y 0)))
+        (shamt XReg (value_regs_get tmp 0))
+        (len_sub_shamt XReg (value_regs_get tmp 1))
+        ;; low part.
+        (low_part1 XReg (rv_sll (value_regs_get x 1) len_sub_shamt))
+        (low_part2 XReg (gen_select_reg (IntCC.Equal) shamt (zero_reg) (zero_reg) low_part1))
+        ;;
+        (low_part3 XReg (rv_srl (value_regs_get x 0) shamt))
+        (low XReg (rv_or low_part2 low_part3))
+        ;;
+        (const64 XReg (imm $I64 64))
+        ;;
+        (high XReg (rv_sra (value_regs_get x 1) shamt))
+        ;;
+        (const_neg_1 XReg (imm $I64 (i64_as_u64 -1)))
+        ;;
+        (high_replacement XReg (gen_select_reg (IntCC.SignedLessThan) (value_regs_get x 1) (zero_reg) const_neg_1 (zero_reg)))
+        (const64 XReg (imm $I64 64))
+        (shamt_128 XReg (rv_andi (value_regs_get y 0) (imm12_const 127))))
+    (value_regs
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high low)
+      (gen_select_reg (IntCC.UnsignedGreaterThanOrEqual) shamt_128 const64 high_replacement high))))
+
+;; SIMD Cases
+;; We don't need to mask or extend anything since it is done by the instruction according to SEW.
+
+(rule 4 (lower (has_type (ty_vec_fits_in_register ty) (sshr x y)))
+  (rv_vsra_vx x (value_regs_get y 0) (unmasked) ty))
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (sshr x (maybe_uextend (uimm5_from_value y)))))
+  (rv_vsra_vi x y (unmasked) ty))
+
+
+;;;; Rules for `rotl` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type (fits_in_64 ty) (rotl x y)))
+  (lower_rotl ty (zext x ty $I64) (value_regs_get y 0)))
+
+(rule 1 (lower (has_type $I128 (rotl x y)))
+  (lower_i128_rotl x y))
+
+;;;; Rules for `rotr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type (fits_in_64 ty) (rotr x y)))
+  (lower_rotr ty (zext x ty $I64) (value_regs_get y 0)))
+
+(rule 1 (lower (has_type $I128 (rotr x y)))
+  (lower_i128_rotr x y))
+
+
+;;;; Rules for `fabs` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (fabs x)))
+  (rv_fabs ty x))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fabs x)))
+  (rv_vfabs_v x (unmasked) ty))
+
+;;;; Rules for `fneg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (fneg x)))
+  (rv_fneg ty x))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fneg x)))
+  (rv_vfneg_v x (unmasked) ty))
+
+;;;; Rules for `fcopysign` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (fcopysign x y)))
+  (rv_fsgnj ty x y))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fcopysign x y)))
+  (rv_vfsgnj_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (fcopysign x (splat y))))
+  (rv_vfsgnj_vf x y (unmasked) ty))
+
+;;;; Rules for `fma` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (fma x y z)))
+  (rv_fmadd ty x y z))
+
+;; (fma x y z) computes x * y + z
+;; vfmacc computes vd[i] = +(vs1[i] * vs2[i]) + vd[i]
+;; We need to reverse the order of the arguments
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fma x y z)))
+  (rv_vfmacc_vv z y x (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (fma (splat x) y z)))
+  (rv_vfmacc_vf z y x (unmasked) ty))
+
+;; vfmsac computes vd[i] = +(vs1[i] * vs2[i]) - vd[i]
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (fma x y (fneg z))))
+  (rv_vfmsac_vv z y x (unmasked) ty))
+
+(rule 6 (lower (has_type (ty_vec_fits_in_register ty) (fma (splat x) y (fneg z))))
+  (rv_vfmsac_vf z y x (unmasked) ty))
+
+;; vfnmacc computes vd[i] = -(vs1[i] * vs2[i]) - vd[i]
+
+(rule 4 (lower (has_type (ty_vec_fits_in_register ty) (fma (fneg x) y (fneg z))))
+  (rv_vfnmacc_vv z y x (unmasked) ty))
+
+(rule 6 (lower (has_type (ty_vec_fits_in_register ty) (fma (fneg (splat x)) y (fneg z))))
+  (rv_vfnmacc_vf z y x (unmasked) ty))
+
+;; vfnmsac computes vd[i] = -(vs1[i] * vs2[i]) + vd[i]
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (fma (fneg x) y z)))
+  (rv_vfnmsac_vv z y x (unmasked) ty))
+
+(rule 5 (lower (has_type (ty_vec_fits_in_register ty) (fma (fneg (splat x)) y z)))
+  (rv_vfnmsac_vf z y x (unmasked) ty))
+
+
+;;;; Rules for `sqrt` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (sqrt x)))
+  (rv_fsqrt ty x))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (sqrt x)))
+  (rv_vfsqrt_v x (unmasked) ty))
+
+;;;; Rules for `AtomicRMW` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule -1
+  ;;
+  (lower
+    (has_type (valid_atomic_transaction ty) (atomic_rmw flags op addr x)))
+  (gen_atomic (get_atomic_rmw_op ty op) addr x (atomic_amo)))
+
+;;; for I8 and I16
+(rule 1
+  (lower
+    (has_type (valid_atomic_transaction (fits_in_16 ty)) (atomic_rmw flags op addr x)))
+  (gen_atomic_rmw_loop op ty addr x))
+
+;;;special for I8 and I16 max min etc.
+;;;because I need uextend or sextend the value.
+(rule 2
+  (lower
+    (has_type (valid_atomic_transaction (fits_in_16 ty)) (atomic_rmw flags (is_atomic_rmw_max_etc op $true) addr x)))
+  (gen_atomic_rmw_loop op ty addr (sext x ty $I64)))
+
+
+(rule 2
+  ;;
+  (lower
+    (has_type (valid_atomic_transaction (fits_in_16 ty)) (atomic_rmw flags (is_atomic_rmw_max_etc op $false) addr x)))
+  ;;
+  (gen_atomic_rmw_loop op ty addr (zext x ty $I64)))
+
+;;;;;  Rules for `AtomicRmwOp.Sub`
+(rule
+  (lower
+    (has_type (valid_atomic_transaction ty) (atomic_rmw flags (AtomicRmwOp.Sub) addr x)))
+  (let
+    ((tmp WritableReg (temp_writable_reg ty))
+     (x2 Reg (rv_neg x)))
+    (gen_atomic (get_atomic_rmw_op ty (AtomicRmwOp.Add)) addr x2 (atomic_amo))))
+
+(decl gen_atomic_rmw_loop (AtomicRmwOp Type XReg XReg) XReg)
+(rule
+  (gen_atomic_rmw_loop op ty addr x)
+  (let
+    ((dst WritableXReg (temp_writable_xreg))
+      (t0 WritableXReg (temp_writable_xreg))
+      (_ Unit (emit (MInst.AtomicRmwLoop (gen_atomic_offset addr ty) op dst ty (gen_atomic_p addr ty) x t0))))
+    (writable_reg_to_reg dst)))
+
+;;;;;  Rules for `AtomicRmwOp.Nand`
+(rule
+  (lower
+    (has_type (valid_atomic_transaction ty) (atomic_rmw flags (AtomicRmwOp.Nand) addr x)))
+    (gen_atomic_rmw_loop (AtomicRmwOp.Nand) ty addr x))
+
+(decl is_atomic_rmw_max_etc (AtomicRmwOp bool) AtomicRmwOp)
+(extern extractor is_atomic_rmw_max_etc is_atomic_rmw_max_etc)
+
+;;;;;  Rules for `atomic load`;;;;;;;;;;;;;;;;;
+(rule
+  (lower (has_type (valid_atomic_transaction ty) (atomic_load flags p)))
+  (gen_atomic_load p ty))
+
+
+;;;;;  Rules for `atomic store`;;;;;;;;;;;;;;;;;
+(rule
+  (lower (atomic_store flags src @ (value_type (valid_atomic_transaction ty)) p))
+  (gen_atomic_store p ty src))
+
+(decl gen_atomic_offset (XReg Type) XReg)
+(rule 1 (gen_atomic_offset p (fits_in_16 ty))
+  (rv_slli (rv_andi p (imm12_const 3)) (imm12_const 3)))
+
+(rule (gen_atomic_offset p _)
+  (zero_reg))
+
+(decl gen_atomic_p (XReg Type) XReg)
+(rule 1 (gen_atomic_p p (fits_in_16 ty))
+  (rv_andi p (imm12_const -4)))
+
+(rule (gen_atomic_p p _)
+  p)
+
+
+;;;;;  Rules for `atomic cas`;;;;;;;;;;;;;;;;;
+(rule
+  (lower (has_type (valid_atomic_transaction ty) (atomic_cas flags p e x)))
+  (let
+    ((t0 WritableReg (temp_writable_reg ty))
+      (dst WritableReg (temp_writable_reg ty))
+      (_ Unit (emit (MInst.AtomicCas (gen_atomic_offset p ty) t0 dst (zext e ty $I64) (gen_atomic_p p ty) x ty))))
+    (writable_reg_to_reg dst)))
+
+;;;;;  Rules for `ireduce`;;;;;;;;;;;;;;;;;
+(rule
+  (lower (has_type ty (ireduce x)))
+  (value_regs_get x 0))
+
+;;;;;  Rules for `fpromote`;;;;;;;;;;;;;;;;;
+(rule (lower (fpromote x))
+  (rv_fcvtds x))
+
+;;;;;  Rules for `fvpromote_low`;;;;;;;;;;;;
+
+(rule (lower (has_type (ty_vec_fits_in_register ty) (fvpromote_low x)))
+  (if-let half_ty (ty_half_width ty))
+  (rv_vfwcvt_f_f_v x (unmasked) (vstate_mf2 half_ty)))
+
+;;;;;  Rules for `fdemote`;;;;;;;;;;;;;;;;;;
+(rule (lower (fdemote x))
+  (rv_fcvtsd x))
+
+;;;;;  Rules for `fvdemote`;;;;;;;;;;;;;;;;;
+
+;; `vfncvt...` leaves the upper bits of the register undefined so
+;; we need to zero them out.
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $F32X4) (fvdemote x)))
+  (if-let zero (i8_to_imm5 0))
+  (let ((narrow VReg (rv_vfncvt_f_f_w x (unmasked) (vstate_mf2 ty)))
+        (mask VReg (gen_vec_mask 0xC)))
+    (rv_vmerge_vim narrow zero mask ty)))
+
+
+;;;;;  Rules for for float arithmetic
+
+
+;;;; Rules for `fadd` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_scalar_float ty) (fadd x y)))
+  (rv_fadd ty x y))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fadd x y)))
+  (rv_vfadd_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (fadd x (splat y))))
+  (rv_vfadd_vf x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (fadd (splat x) y)))
+  (rv_vfadd_vf y x (unmasked) ty))
+
+
+;;;; Rules for `fsub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (fsub x y)))
+  (rv_fsub ty x y))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fsub x y)))
+  (rv_vfsub_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (fsub x (splat y))))
+  (rv_vfsub_vf x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (fsub (splat x) y)))
+  (rv_vfrsub_vf y x (unmasked) ty))
+
+;;;; Rules for `fmul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (fmul x y)))
+  (rv_fmul ty x y))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fmul x y)))
+  (rv_vfmul_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (fmul x (splat y))))
+  (rv_vfmul_vf x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (fmul (splat x) y)))
+  (rv_vfmul_vf y x (unmasked) ty))
+
+
+;;;; Rules for `fdiv` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (fdiv x y)))
+  (rv_fdiv ty x y))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fdiv x y)))
+  (rv_vfdiv_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (fdiv x (splat y))))
+  (rv_vfdiv_vf x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (fdiv (splat x) y)))
+  (rv_vfrdiv_vf y x (unmasked) ty))
+
+;;;; Rules for `fmin` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_scalar_float ty) (fmin x y)))
+  (gen_float_select (FloatSelectOP.Min) x y ty))
+
+;; vfmin does almost the right thing, but it does not handle NaN's correctly.
+;; We should return a NaN if any of the inputs is a NaN, but vfmin returns the
+;; number input instead.
+;;
+;; TODO: We can improve this by using a masked `fmin` instruction that modifies
+;; the canonical nan register. That way we could avoid the `vmerge.vv` instruction.
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fmin x y)))
+  (let ((is_not_nan VReg (gen_fcmp_mask ty (FloatCC.Ordered) x y))
+        (nan XReg (imm $I64 (canonical_nan_u64 (lane_type ty))))
+        (vec_nan VReg (rv_vmv_vx nan ty))
+        (min VReg (rv_vfmin_vv x y (unmasked) ty)))
+    (rv_vmerge_vvm vec_nan min is_not_nan ty)))
+
+;;;; Rules for `fmax` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_scalar_float ty) (fmax x y)))
+  (gen_float_select (FloatSelectOP.Max) x y ty))
+
+;; vfmax does almost the right thing, but it does not handle NaN's correctly.
+;; We should return a NaN if any of the inputs is a NaN, but vfmax returns the
+;; number input instead.
+;;
+;; TODO: We can improve this by using a masked `fmax` instruction that modifies
+;; the canonical nan register. That way we could avoid the `vmerge.vv` instruction.
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (fmax x y)))
+  (let ((is_not_nan VReg (gen_fcmp_mask ty (FloatCC.Ordered) x y))
+        (nan XReg (imm $I64 (canonical_nan_u64 (lane_type ty))))
+        (vec_nan VReg (rv_vmv_vx nan ty))
+        (max VReg (rv_vfmax_vv x y (unmasked) ty)))
+    (rv_vmerge_vvm vec_nan max is_not_nan ty)))
+
+;;;;;  Rules for `stack_addr`;;;;;;;;;
+(rule
+  (lower (stack_addr ss offset))
+  (gen_stack_addr ss offset))
+
+;;;;;  Rules for `is_null`;;;;;;;;;
+
+;; Null references are represented by the constant value `0`.
+(rule (lower (is_null v))
+  (rv_seqz v))
+
+;;;;;  Rules for `is_invalid`;;;;;;;;;
+
+;; Invalid references are represented by the constant value `-1`.
+(rule (lower (is_invalid v))
+  (rv_seqz (rv_addi v (imm12_const 1))))
+
+;;;;;  Rules for `select`;;;;;;;;;
+(rule
+  (lower (has_type ty (select c @ (value_type cty) x y)))
+  (gen_select ty (truthy_to_reg cty (normalize_cmp_value cty c (ExtendOp.Zero))) x y))
+
+(rule 1
+  (lower (has_type (fits_in_64 ty) (select (icmp cc a b @ (value_type (fits_in_64 in_ty))) x y)))
+  (let ((a XReg (truthy_to_reg in_ty (normalize_cmp_value in_ty a (intcc_to_extend_op cc))))
+        (b XReg (truthy_to_reg in_ty (normalize_cmp_value in_ty b (intcc_to_extend_op cc)))))
+    (gen_select_reg cc a b x y)))
+
+;;;;;  Rules for `bitselect`;;;;;;;;;
+
+;; Do a (c & x) | (~c & y) operation.
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (bitselect c x y)))
+  (let ((tmp_x XReg (rv_and c x))
+        (c_inverse XReg (rv_not c))
+        (tmp_y XReg (rv_and c_inverse y)))
+    (rv_or tmp_x tmp_y)))
+
+;; For vectors, we also do the same operation.
+;; We can technically use any type in the bitwise operations, but prefer
+;; using the type of the inputs so that we avoid emitting unnecessary
+;; `vsetvl` instructions. It's likeley that the vector unit is already
+;; configured for that type.
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (bitselect c x y)))
+  (let ((tmp_x VReg (rv_vand_vv c x (unmasked) ty))
+        (c_inverse VReg (rv_vnot_v c (unmasked) ty))
+        (tmp_y VReg (rv_vand_vv c_inverse y (unmasked) ty)))
+    (rv_vor_vv tmp_x tmp_y (unmasked) ty)))
+
+;; Special case for bitselects with cmp's as an input.
+;;
+;; This allows us to skip the mask expansion step and use the more efficient
+;; vmerge.vvm instruction.
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (bitselect (icmp cc a @ (value_type (ty_vec_fits_in_register cmp_ty)) b) x y)))
+  (let ((mask VReg (gen_icmp_mask cmp_ty cc a b)))
+    (rv_vmerge_vvm y x mask ty)))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (bitselect (fcmp cc a @ (value_type (ty_vec_fits_in_register cmp_ty)) b) x y)))
+  (let ((mask VReg (gen_fcmp_mask cmp_ty cc a b)))
+    (rv_vmerge_vvm y x mask ty)))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (bitselect (bitcast _ (fcmp cc a @ (value_type (ty_vec_fits_in_register cmp_ty)) b)) x y)))
+  (let ((mask VReg (gen_fcmp_mask cmp_ty cc a b)))
+    (rv_vmerge_vvm y x mask ty)))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (bitselect (bitcast _ (icmp cc a @ (value_type (ty_vec_fits_in_register cmp_ty)) b)) x y)))
+  (let ((mask VReg (gen_icmp_mask cmp_ty cc a b)))
+    (rv_vmerge_vvm y x mask ty)))
+
+
+;;;;;  Rules for `isplit`;;;;;;;;;
+(rule
+  (lower (isplit x))
+  (let
+    ((t1 XReg (value_regs_get x 0))
+      (t2 XReg (value_regs_get x 1)))
+    (output_pair t1 t2)))
+
+;;;;;  Rules for `iconcat`;;;;;;;;;
+(rule
+  (lower (has_type $I128 (iconcat x y)))
+  (let
+    ((t1 XReg x)
+      (t2 XReg y))
+    (value_regs t1 t2)))
+
+
+;;;;;  Rules for `smax`;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_int ty)  (smax x y)))
+  (gen_int_select ty (IntSelectOP.Smax) (ext_int_if_need $true x ty) (ext_int_if_need $true y ty)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (smax x y)))
+  (rv_vmax_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (smax x (splat y))))
+  (rv_vmax_vx x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (smax (splat x) y)))
+  (rv_vmax_vx y x (unmasked) ty))
+
+;;;;;  Rules for `smin`;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_int ty)  (smin x y)))
+  (gen_int_select ty (IntSelectOP.Smin) (ext_int_if_need $true x ty) (ext_int_if_need $true y ty)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (smin x y)))
+  (rv_vmin_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (smin x (splat y))))
+  (rv_vmin_vx x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (smin (splat x) y)))
+  (rv_vmin_vx y x (unmasked) ty))
+
+;;;;;  Rules for `umax`;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_int ty)  (umax x y)))
+  (gen_int_select ty (IntSelectOP.Umax) (ext_int_if_need $false x ty) (ext_int_if_need $false y ty)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (umax x y)))
+  (rv_vmaxu_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (umax x (splat y))))
+  (rv_vmaxu_vx x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (umax (splat x) y)))
+  (rv_vmaxu_vx y x (unmasked) ty))
+
+;;;;;  Rules for `umin`;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_int ty) (umin x y)))
+  (gen_int_select ty (IntSelectOP.Umin) (ext_int_if_need $false x ty) (ext_int_if_need $false y ty)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (umin x y)))
+  (rv_vminu_vv x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (umin x (splat y))))
+  (rv_vminu_vx x y (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (umin (splat x) y)))
+  (rv_vminu_vx y x (unmasked) ty))
+
+
+;;;;;  Rules for `debugtrap`;;;;;;;;;
+(rule
+  (lower (debugtrap))
+  (side_effect (SideEffectNoResult.Inst (MInst.EBreak))))
+
+;;;;;  Rules for `fence`;;;;;;;;;
+(rule
+  (lower (fence))
+  (side_effect (SideEffectNoResult.Inst (MInst.Fence 15 15))))
+
+;;;;;  Rules for `trap`;;;;;;;;;
+(rule
+  (lower (trap code))
+  (udf code))
+
+;;;;;  Rules for `resumable_trap`;;;;;;;;;
+(rule
+  (lower (resumable_trap code))
+  (udf code))
+
+;;;;;  Rules for `uload8`;;;;;;;;;
+(rule
+  (lower (uload8 flags p @ (value_type (ty_addr64 _)) offset))
+  (gen_load p offset (int_load_op $false 8) flags $I64))
+;;;;;  Rules for `sload8`;;;;;;;;;
+(rule
+  (lower (sload8 flags p @ (value_type (ty_addr64 _)) offset))
+  (gen_load p offset (int_load_op $true 8) flags $I64))
+;;;;;  Rules for `uload16`;;;;;;;;;
+(rule
+  (lower (uload16 flags p @ (value_type (ty_addr64 _)) offset))
+  (gen_load p offset (int_load_op $false 16) flags $I64))
+
+;;;;;  Rules for `iload16`;;;;;;;;;
+(rule
+  (lower (sload16 flags p @ (value_type (ty_addr64 _)) offset))
+  (gen_load p offset (int_load_op $true 16) flags $I64))
+
+;;;;;  Rules for `uload32`;;;;;;;;;
+(rule
+  (lower (uload32 flags p @ (value_type (ty_addr64 _)) offset))
+  (gen_load p offset (int_load_op $false 32) flags $I64))
+
+;;;;;  Rules for `iload16`;;;;;;;;;
+(rule
+  (lower (sload32 flags p @ (value_type (ty_addr64 _)) offset))
+  (gen_load p offset (int_load_op $true 32) flags $I64))
+
+(rule
+  (lower (has_type ty (load flags p @ (value_type (ty_addr64 _)) offset)))
+  (gen_load p offset (load_op ty) flags ty)
+)
+;;;; for I128
+(rule 1
+  (lower (has_type $I128 (load flags p @ (value_type (ty_addr64 _)) offset)))
+  (gen_load_128 p offset flags))
+
+(rule 2
+  (lower (has_type (ty_vec_fits_in_register ty) (load flags p @ (value_type (ty_addr64 _)) offset)))
+  (let ((eew VecElementWidth (element_width_from_type ty)))
+    (vec_load eew (VecAMode.UnitStride (gen_amode p offset $I64)) flags (unmasked) ty)))
+
+;;;;;  Rules for Load + Extend Combos ;;;;;;;;;
+
+;; These rules cover the special loads that load a 64bit value and do some sort of extension.
+;; We don't have any special instructions to do this, so just load the 64 bits as a vector, and
+;; do a SEW/2 extension. This only reads half width elements from the source vector register
+;; extends it, and writes the back the full register.
+
+(decl gen_load64_extend (Type ExtendOp MemFlags XReg Offset32) VReg)
+
+(rule (gen_load64_extend ty (ExtendOp.Signed) flags addr offset)
+  (let ((eew VecElementWidth (element_width_from_type $I64))
+        (load_state VState (vstate_from_type $I64))
+        (loaded VReg (vec_load eew (VecAMode.UnitStride (gen_amode addr offset $I64)) flags (unmasked) load_state)))
+    (rv_vsext_vf2 loaded (unmasked) ty)))
+
+(rule (gen_load64_extend ty (ExtendOp.Zero) flags addr offset)
+  (let ((eew VecElementWidth (element_width_from_type $I64))
+        (load_state VState (vstate_from_type $I64))
+        (loaded VReg (vec_load eew (VecAMode.UnitStride (gen_amode addr offset $I64)) flags (unmasked) load_state)))
+    (rv_vzext_vf2 loaded (unmasked) ty)))
+
+;;;;;  Rules for `uload8x8`;;;;;;;;;;
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $I16X8) (uload8x8 flags addr @ (value_type (ty_addr64 _)) offset)))
+  (gen_load64_extend ty (ExtendOp.Zero) flags addr offset))
+
+;;;;;  Rules for `uload16x4`;;;;;;;;;
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $I32X4) (uload16x4 flags addr @ (value_type (ty_addr64 _)) offset)))
+  (gen_load64_extend ty (ExtendOp.Zero) flags addr offset))
+
+;;;;;  Rules for `uload32x2`;;;;;;;;;
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $I64X2) (uload32x2 flags addr @ (value_type (ty_addr64 _)) offset)))
+  (gen_load64_extend ty (ExtendOp.Zero) flags addr offset))
+
+;;;;;  Rules for `sload8x8`;;;;;;;;;;
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $I16X8) (sload8x8 flags addr @ (value_type (ty_addr64 _)) offset)))
+  (gen_load64_extend ty (ExtendOp.Signed) flags addr offset))
+
+;;;;;  Rules for `sload16x4`;;;;;;;;;
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $I32X4) (sload16x4 flags addr @ (value_type (ty_addr64 _)) offset)))
+  (gen_load64_extend ty (ExtendOp.Signed) flags addr offset))
+
+;;;;;  Rules for `sload32x2`;;;;;;;;;
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $I64X2) (sload32x2 flags addr @ (value_type (ty_addr64 _)) offset)))
+  (gen_load64_extend ty (ExtendOp.Signed) flags addr offset))
+
+;;;;;  Rules for `istore8`;;;;;;;;;
+(rule
+  (lower (istore8 flags x p @ (value_type (ty_addr64 _)) offset))
+  (gen_store p offset (StoreOP.Sb) flags x))
+;;;;;  Rules for `istore16`;;;;;;;;;
+(rule
+  (lower (istore16 flags x p @ (value_type (ty_addr64 _)) offset))
+  (gen_store p offset (StoreOP.Sh) flags x))
+
+;;;;;  Rules for `istore32`;;;;;;;;;
+(rule
+  (lower (istore32 flags x p @ (value_type (ty_addr64 _)) offset))
+  (gen_store p offset (StoreOP.Sw) flags x))
+
+;;;;;  Rules for `store`;;;;;;;;;
+(rule
+  (lower (store flags x @ (value_type ty) p @ (value_type (ty_addr64 _)) offset))
+  (gen_store p offset (store_op ty) flags x))
+
+;;; special for I128
+(rule 1
+  (lower (store flags x @ (value_type $I128 ) p @ (value_type (ty_addr64 _)) offset))
+  (gen_store_128 p offset flags x))
+
+(rule 2
+  (lower (store flags x @ (value_type (ty_vec_fits_in_register ty)) p @ (value_type (ty_addr64 _)) offset))
+  (let ((eew VecElementWidth (element_width_from_type ty)))
+    (vec_store eew (VecAMode.UnitStride (gen_amode p offset $I64)) x flags (unmasked) ty)))
+
+(decl gen_icmp (IntCC ValueRegs ValueRegs Type) XReg)
+(rule
+  (gen_icmp cc x y ty)
+  (let
+    ((result WritableXReg (temp_writable_xreg))
+      (_ Unit (emit (MInst.Icmp cc result x y ty))))
+    result))
+
+;;;;;  Rules for `icmp`;;;;;;;;;
+(rule 0 (lower (icmp cc x @ (value_type (ty_int ty)) y))
+  (lower_icmp cc x y ty))
+
+(rule 1 (lower (icmp cc x @ (value_type (ty_vec_fits_in_register ty)) y))
+  (gen_expand_mask ty (gen_icmp_mask ty cc x y)))
+
+
+;;;;;  Rules for `fcmp`;;;;;;;;;
+(rule 0 (lower (fcmp cc x @ (value_type (ty_scalar_float ty)) y))
+  (cmp_value (emit_fcmp cc ty x y)))
+
+(rule 1 (lower (fcmp cc x @ (value_type (ty_vec_fits_in_register ty)) y))
+  (gen_expand_mask ty (gen_fcmp_mask ty cc x y)))
+
+;;;;;  Rules for `func_addr`;;;;;;;;;
+(rule
+  (lower (func_addr (func_ref_data _ name _)))
+  (load_ext_name name 0))
+
+;;;;;  Rules for `fcvt_to_uint`;;;;;;;;;
+(rule
+  (lower (has_type to (fcvt_to_uint v @ (value_type from))))
+  (gen_fcvt_int $false v $false from to))
+
+;;;;;  Rules for `fcvt_to_sint`;;;;;;;;;
+(rule 0 (lower (has_type to (fcvt_to_sint v @ (value_type (ty_scalar_float from)))))
+  (gen_fcvt_int $false v $true from to))
+
+;;;;;  Rules for `fcvt_to_sint_sat`;;;;;;;;;
+(rule 0 (lower (has_type to (fcvt_to_sint_sat v @ (value_type (ty_scalar_float from)))))
+  (gen_fcvt_int $true v $true from to))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register _) (fcvt_to_sint_sat v @ (value_type from_ty))))
+  (if-let zero (i8_to_imm5 0))
+  (let ((is_nan VReg (rv_vmfne_vv v v (unmasked) from_ty))
+        (cvt VReg (rv_vfcvt_rtz_x_f_v v (unmasked) from_ty)))
+    (rv_vmerge_vim cvt zero is_nan from_ty)))
+
+;;;;;  Rules for `fcvt_to_uint_sat`;;;;;;;;;
+(rule 0 (lower (has_type to (fcvt_to_uint_sat v @ (value_type (ty_scalar_float from)))))
+  (gen_fcvt_int $true v $false from to))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register _) (fcvt_to_uint_sat v @ (value_type from_ty))))
+  (if-let zero (i8_to_imm5 0))
+  (let ((is_nan VReg (rv_vmfne_vv v v (unmasked) from_ty))
+        (cvt VReg (rv_vfcvt_rtz_xu_f_v v (unmasked) from_ty)))
+    (rv_vmerge_vim cvt zero is_nan from_ty)))
+
+;;;;;  Rules for `fcvt_from_sint`;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float to) (fcvt_from_sint v @ (value_type from_ty))))
+  (let ((float_op FpuOPRR (int_convert_2_float_op from_ty $true to))
+        (value XReg (normalize_fcvt_from_int v from_ty (ExtendOp.Signed))))
+    (fpu_rr float_op to value)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register _) (fcvt_from_sint v @ (value_type from_ty))))
+  (rv_vfcvt_f_x_v v (unmasked) from_ty))
+
+;;;;;  Rules for `fcvt_from_uint`;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float to) (fcvt_from_uint v @ (value_type from_ty))))
+  (let ((float_op FpuOPRR (int_convert_2_float_op from_ty $false to))
+        (value XReg (normalize_fcvt_from_int v from_ty (ExtendOp.Zero))))
+    (fpu_rr float_op to value)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register _) (fcvt_from_uint v @ (value_type from_ty))))
+  (rv_vfcvt_f_xu_v v (unmasked) from_ty))
+
+;;;;;  Rules for `symbol_value`;;;;;;;;;
+(rule
+   (lower (symbol_value (symbol_value_data name _ offset)))
+   (load_ext_name name offset))
+
+;;;;;  Rules for `bitcast`;;;;;;;;;
+(rule
+   (lower (has_type out_ty (bitcast _ v @ (value_type in_ty))))
+   (gen_bitcast v in_ty out_ty))
+
+;;;;;  Rules for `ceil`;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (ceil x)))
+  (gen_float_round (FloatRoundOP.Ceil) x ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (ceil x)))
+  (gen_vec_round x (FRM.RUP) ty))
+
+;;;;;  Rules for `floor`;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (floor x)))
+  (gen_float_round (FloatRoundOP.Floor) x ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (floor x)))
+  (gen_vec_round x (FRM.RDN) ty))
+
+;;;;;  Rules for `trunc`;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (trunc x)))
+  (gen_float_round (FloatRoundOP.Trunc) x ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (trunc x)))
+  (gen_vec_round x (FRM.RTZ) ty))
+
+;;;;;  Rules for `nearest`;;;;;;;;;
+(rule 0 (lower (has_type (ty_scalar_float ty) (nearest x)))
+  (gen_float_round (FloatRoundOP.Nearest) x ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (nearest x)))
+  (gen_vec_round x (FRM.RNE) ty))
+
+
+;;;;;  Rules for `select_spectre_guard`;;;;;;;;;
+
+;; SelectSpectreGuard is equivalent to Select, but we should not use a branch based
+;; lowering for it. Instead we use a conditional move based lowering.
+;;
+;; We don't have cmov's in RISC-V either, but we can emulate those using bitwise
+;; operations, which is what we do below.
+(rule (lower (has_type ty (select_spectre_guard cmp @ (value_type cmp_ty) x @ (value_type arg_ty) y)))
+  (let (;; Build a mask that is 0 or -1 depending on the input comparision value.
+        ;; `lower_bmask` handles normalizing the input.
+        (mask ValueRegs (lower_bmask arg_ty cmp_ty cmp))
+        ;; Using the mask above we can select either `x` or `y` by
+        ;; performing a bitwise `and` on both sides and then merging them
+        ;; together. We know that only the bits of one of the sides will be selected.
+        ;; TODO: We can use `andn` here if we have `Zbb`
+        (lhs ValueRegs (gen_and arg_ty x mask))
+        (rhs ValueRegs (gen_and arg_ty y (gen_bnot arg_ty mask))))
+    (gen_or arg_ty lhs rhs)))
+
+;;;;;  Rules for `bmask`;;;;;;;;;
+(rule
+  (lower (has_type oty (bmask x @ (value_type ity))))
+  (lower_bmask oty ity x))
+
+;; N.B.: the Ret itself is generated by the ABI.
+(rule (lower (return args))
+      (lower_return args))
+
+;;; Rules for `get_{frame,stack}_pointer` and `get_return_address` ;;;;;;;;;;;;;
+
+(rule (lower (get_frame_pointer))
+  (gen_mov_from_preg (fp_reg)))
+
+(rule (lower (get_stack_pointer))
+  (gen_mov_from_preg (sp_reg)))
+
+(rule (lower (get_return_address))
+  (load_ra))
+
+;;; Rules for `iabs` ;;;;;;;;;;;;;
+
+;; I64 and lower
+;; Generate the following code:
+;;   sext.{b,h,w} a0, a0
+;;   neg a1, a0
+;;   max a0, a0, a1
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (iabs x)))
+  (let ((extended XReg (sext x ty $I64))
+        (negated XReg (rv_neg extended)))
+    (max $I64 extended negated)))
+
+;; For vectors we generate the same code, but with vector instructions
+;; we can skip the sign extension, since the vector unit will only process
+;; Element Sized chunks.
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (iabs x)))
+  (let ((negated VReg (rv_vneg_v x (unmasked) ty)))
+    (rv_vmax_vv x negated (unmasked) ty)))
+
+;;;; Rules for calls ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(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 `return_call` and `return_call_indirect` ;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (return_call (func_ref_data sig_ref extname dist) args))
+      (gen_return_call sig_ref extname dist args))
+
+(rule (lower (return_call_indirect sig_ref callee args))
+      (gen_return_call_indirect sig_ref callee args))
+
+
+;;;; Rules for `extractlane` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (extractlane x @ (value_type ty) (u8_from_uimm8 idx)))
+  (gen_extractlane ty x idx))
+
+;;;; Rules for `insertlane` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; We can insert a lane by using a masked splat from an X register.
+;; Build a mask that is only enabled in the lane we want to insert.
+;; Then use a masked splat (vmerge) to insert the value.
+(rule 0 (lower (insertlane vec @ (value_type (ty_vec_fits_in_register ty))
+                           val @ (value_type (ty_int _))
+                           (u8_from_uimm8 lane)))
+  (let ((mask VReg (gen_vec_mask (u64_shl 1 lane))))
+    (rv_vmerge_vxm vec val mask ty)))
+
+;; Similar to above, but using the float variants of the instructions.
+(rule 1 (lower (insertlane vec @ (value_type (ty_vec_fits_in_register ty))
+                           val @ (value_type (ty_scalar_float _))
+                           (u8_from_uimm8 lane)))
+  (let ((mask VReg (gen_vec_mask (u64_shl 1 lane))))
+    (rv_vfmerge_vfm vec val mask ty)))
+
+;; If we are inserting from an Imm5 const we can use the immediate
+;; variant of vmerge.
+(rule 2 (lower (insertlane vec @ (value_type (ty_vec_fits_in_register ty))
+                           (i64_from_iconst (imm5_from_i64 imm))
+                           (u8_from_uimm8 lane)))
+  (let ((mask VReg (gen_vec_mask (u64_shl 1 lane))))
+    (rv_vmerge_vim vec imm mask ty)))
+
+;;;; Rules for `splat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type ty (splat n @ (value_type (ty_scalar_float _)))))
+  (rv_vfmv_vf n ty))
+
+(rule 1 (lower (has_type ty (splat n @ (value_type (ty_int_ref_scalar_64 _)))))
+  (rv_vmv_vx n ty))
+
+(rule 2 (lower (has_type ty (splat (iconst (u64_from_imm64 (imm5_from_u64 imm))))))
+  (rv_vmv_vi imm ty))
+
+;; TODO: We can splat out more patterns by using for example a vmv.v.i i8x16 for
+;; a i64x2 const with a compatible bit pattern. The AArch64 Backend does something
+;; similar in its splat rules.
+;; TODO: Look through bitcasts when splatting out registers. We can use
+;; `vmv.v.x` in a `(splat.f32x4 (bitcast.f32 val))`. And vice versa for integers.
+
+;;;; Rules for `uadd_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register ty) (uadd_sat x y)))
+  (rv_vsaddu_vv x y (unmasked) ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (uadd_sat x (splat y))))
+  (rv_vsaddu_vx x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (uadd_sat (splat x) y)))
+  (rv_vsaddu_vx y x (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (uadd_sat x (replicated_imm5 y))))
+  (rv_vsaddu_vi x y (unmasked) ty))
+
+(rule 4 (lower (has_type (ty_vec_fits_in_register ty) (uadd_sat (replicated_imm5 x) y)))
+  (rv_vsaddu_vi y x (unmasked) ty))
+
+;;;; Rules for `sadd_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register ty) (sadd_sat x y)))
+  (rv_vsadd_vv x y (unmasked) ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (sadd_sat x (splat y))))
+  (rv_vsadd_vx x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (sadd_sat (splat x) y)))
+  (rv_vsadd_vx y x (unmasked) ty))
+
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (sadd_sat x (replicated_imm5 y))))
+  (rv_vsadd_vi x y (unmasked) ty))
+
+(rule 4 (lower (has_type (ty_vec_fits_in_register ty) (sadd_sat (replicated_imm5 x) y)))
+  (rv_vsadd_vi y x (unmasked) ty))
+
+;;;; Rules for `usub_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register ty) (usub_sat x y)))
+  (rv_vssubu_vv x y (unmasked) ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (usub_sat x (splat y))))
+  (rv_vssubu_vx x y (unmasked) ty))
+
+;;;; Rules for `ssub_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register ty) (ssub_sat x y)))
+  (rv_vssub_vv x y (unmasked) ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (ssub_sat x (splat y))))
+  (rv_vssub_vx x y (unmasked) ty))
+
+;;;; Rules for `vall_true` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; Here we do a Vector Reduce operation. Get the unsigned minimum value of any
+;; lane in the vector. The fixed input to the reduce operation is a 1.
+;; This way, if any lane is 0, the result will be 0. Otherwise, the result will
+;; be a 1.
+;; The reduce operation leaves the result in the lowest lane, we then move it
+;; into the destination X register.
+(rule (lower (vall_true x @ (value_type (ty_vec_fits_in_register ty))))
+  (if-let one (i8_to_imm5 1))
+  ;; We don't need to broadcast the immediate into all lanes, only into lane 0.
+  ;; I did it this way since it uses one less instruction than with a vmv.s.x.
+  (let ((fixed VReg (rv_vmv_vi one ty))
+        (min VReg (rv_vredminu_vs x fixed (unmasked) ty)))
+    (rv_vmv_xs min ty)))
+
+
+;;;; Rules for `vany_true` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; Here we do a Vector Reduce operation. Get the unsigned maximum value of the
+;; input vector register. Move the max to an X register, and do a `snez` on it
+;; to ensure its either 1 or 0.
+(rule (lower (vany_true x @ (value_type (ty_vec_fits_in_register ty))))
+  (let ((max VReg (rv_vredmaxu_vs x x (unmasked) ty))
+        (x_max XReg (rv_vmv_xs max ty)))
+    (rv_snez x_max)))
+
+
+;;;; Rules for `vhigh_bits` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; To check if the MSB of a lane is set, we do a `vmslt` with zero, this sets
+;; the mask bit to 1 if the value is negative (MSB 1) and 0 if not. We can then
+;; just move that mask to an X Register.
+;;
+;; We must ensure that the move to the X register has a SEW with enough bits
+;; to hold the full mask. Additionally, in some cases (e.g. i64x2) we are going
+;; to read some tail bits. These are undefined, so we need to further mask them
+;; off.
+(rule (lower (vhigh_bits x @ (value_type (ty_vec_fits_in_register ty))))
+  (let ((mask VReg (rv_vmslt_vx x (zero_reg) (unmasked) ty))
+        ;; Here we only need I64X1, but emit an AVL of 2 since it
+        ;; saves one vector state change in the case of I64X2.
+        ;;
+        ;; TODO: For types that have more lanes than element bits, we can
+        ;; use the original type as a VState and avoid a state change.
+        (x_mask XReg (rv_vmv_xs mask (vstate_from_type $I64X2))))
+    (gen_andi x_mask (ty_lane_mask ty))))
+
+;;;; Rules for `swizzle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register ty) (swizzle x y)))
+  (rv_vrgather_vv x y (unmasked) ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (swizzle x (splat y))))
+  (rv_vrgather_vx x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (swizzle x (replicated_uimm5 y))))
+  (rv_vrgather_vi x y (unmasked) ty))
+
+;;;; Rules for `shuffle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; Use a vrgather to load all 0-15 lanes from x. And then modify the mask to load all
+;; 16-31 lanes from y. Finally, use a vor to combine the two vectors.
+;;
+;; vrgather will insert a 0 for lanes that are out of bounds, so we can let it load
+;; negative and out of bounds indexes.
+(rule (lower (has_type (ty_vec_fits_in_register ty @ $I8X16) (shuffle x y (vconst_from_immediate mask))))
+  (if-let neg16 (i8_to_imm5 -16))
+  (let ((x_mask VReg (gen_constant ty mask))
+        (x_lanes VReg (rv_vrgather_vv x x_mask (unmasked) ty))
+        (y_mask VReg (rv_vadd_vi x_mask neg16 (unmasked) ty))
+        (y_lanes VReg (rv_vrgather_vv y y_mask (unmasked) ty)))
+    (rv_vor_vv x_lanes y_lanes (unmasked) ty)))
+
+;;;; Rules for `swiden_high` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; Slide down half the vector, and do a signed extension.
+(rule 0 (lower (has_type (ty_vec_fits_in_register out_ty) (swiden_high x @ (value_type in_ty))))
+  (rv_vsext_vf2 (gen_slidedown_half in_ty x) (unmasked) out_ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register out_ty) (swiden_high (swiden_high x @ (value_type in_ty)))))
+  (if-let (uimm5_from_u64 amt) (u64_sub (ty_lane_count in_ty) (ty_lane_count out_ty)))
+  (rv_vsext_vf4 (rv_vslidedown_vi x amt (unmasked) in_ty) (unmasked) out_ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register out_ty) (swiden_high (swiden_high (swiden_high x @ (value_type in_ty))))))
+  (if-let (uimm5_from_u64 amt) (u64_sub (ty_lane_count in_ty) (ty_lane_count out_ty)))
+  (rv_vsext_vf8 (rv_vslidedown_vi x amt (unmasked) in_ty) (unmasked) out_ty))
+
+;;;; Rules for `uwiden_high` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; Slide down half the vector, and do a zero extension.
+(rule 0 (lower (has_type (ty_vec_fits_in_register out_ty) (uwiden_high x @ (value_type in_ty))))
+  (rv_vzext_vf2 (gen_slidedown_half in_ty x) (unmasked) out_ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register out_ty) (uwiden_high (uwiden_high x @ (value_type in_ty)))))
+  (if-let (uimm5_from_u64 amt) (u64_sub (ty_lane_count in_ty) (ty_lane_count out_ty)))
+  (rv_vzext_vf4 (rv_vslidedown_vi x amt (unmasked) in_ty) (unmasked) out_ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register out_ty) (uwiden_high (uwiden_high (uwiden_high x @ (value_type in_ty))))))
+  (if-let (uimm5_from_u64 amt) (u64_sub (ty_lane_count in_ty) (ty_lane_count out_ty)))
+  (rv_vzext_vf8 (rv_vslidedown_vi x amt (unmasked) in_ty) (unmasked) out_ty))
+
+;;;; Rules for `swiden_low` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register out_ty) (swiden_low x)))
+  (rv_vsext_vf2 x (unmasked) out_ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register out_ty) (swiden_low (swiden_low x))))
+  (rv_vsext_vf4 x (unmasked) out_ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register out_ty) (swiden_low (swiden_low (swiden_low x)))))
+  (rv_vsext_vf8 x (unmasked) out_ty))
+
+;;;; Rules for `uwiden_low` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register out_ty) (uwiden_low x)))
+  (rv_vzext_vf2 x (unmasked) out_ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register out_ty) (uwiden_low (uwiden_low x))))
+  (rv_vzext_vf4 x (unmasked) out_ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register out_ty) (uwiden_low (uwiden_low (uwiden_low x)))))
+  (rv_vzext_vf8 x (unmasked) out_ty))
+
+;;;; Rules for `iadd_pairwise` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; We don't have a dedicated instruction for this, rearrange the register elements
+;; and use a vadd.
+;;
+;; We do this by building two masks, one for the even elements and one for the odd
+;; elements. Using vcompress we can extract the elements and group them together.
+;;
+;; This is likely not the optimal way of doing this. LLVM does this using a bunch
+;; of vrgathers (See: https://godbolt.org/z/jq8Wj8WG4), that doesen't seem to be
+;; too much better than this.
+;;
+;; However V8 does something better. They use 2 vcompresses using LMUL2, that means
+;; that they can do the whole thing in 3 instructions (2 vcompress + vadd). We don't
+;; support LMUL > 1, so we can't do that.
+(rule (lower (has_type (ty_vec_fits_in_register ty) (iadd_pairwise x y)))
+  (if-let half_size (u64_to_uimm5 (u64_udiv (ty_lane_count ty) 2)))
+  (let ((odd_mask  VReg (gen_vec_mask 0x5555555555555555))
+        (lhs_lo VReg (rv_vcompress_vm x odd_mask ty))
+        (lhs_hi VReg (rv_vcompress_vm y odd_mask ty))
+        (lhs VReg (rv_vslideup_vvi lhs_lo lhs_hi half_size (unmasked) ty))
+
+        (even_mask VReg (gen_vec_mask 0xAAAAAAAAAAAAAAAA))
+        (rhs_lo VReg (rv_vcompress_vm x even_mask ty))
+        (rhs_hi VReg (rv_vcompress_vm y even_mask ty))
+        (rhs VReg (rv_vslideup_vvi rhs_lo rhs_hi half_size (unmasked) ty)))
+    (rv_vadd_vv lhs rhs (unmasked) ty)))
+
+;;;; Rules for `avg_round` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; `avg_round` computes the unsigned average with rounding: a := (x + y + 1) // 2
+;;
+;; See Section "2–5 Average of Two Integers" of the Hacker's Delight book
+;;
+;; The floor average of two integers without overflow can be computed as:
+;;     t = (x & y) + ((x ^ y) >> 1)
+;;
+;; The right shift should be a logical shift if the integers are unsigned.
+;;
+;; We are however interested in the ceiling average (x + y + 1). For that
+;; we use a special rounding mode in the right shift instruction.
+;;
+;; For the right shift instruction we use `vssrl` which is a Scaling Shift
+;; Right Logical instruction using the `vxrm` fixed-point rouding mode. The
+;; default rounding mode is `rnu` (round-to-nearest-up (add +0.5 LSB)).
+;; Which is coincidentally the rounding mode we want for `avg_round`.
+(rule (lower (has_type (ty_vec_fits_in_register ty) (avg_round x y)))
+  (if-let one (u64_to_uimm5 1))
+  (let ((lhs VReg (rv_vand_vv x y (unmasked) ty))
+        (xor  VReg (rv_vxor_vv x y (unmasked) ty))
+        (rhs VReg (rv_vssrl_vi xor one (unmasked) ty)))
+    (rv_vadd_vv lhs rhs (unmasked) ty)))
+
+;;;; Rules for `scalar_to_vector` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (scalar_to_vector x)))
+  (if (ty_vector_not_float ty))
+  (let ((zero VReg (rv_vmv_vx (zero_reg) ty))
+        (mask VReg (gen_vec_mask 1)))
+    (rv_vmerge_vxm zero x mask ty)))
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register ty) (scalar_to_vector x)))
+  (if (ty_vector_float ty))
+  (let ((zero VReg (rv_vmv_vx (zero_reg) ty))
+        (elem VReg (rv_vfmv_sf x ty))
+        (mask VReg (gen_vec_mask 1)))
+    (rv_vmerge_vvm zero elem mask ty)))
+
+;;;; Rules for `sqmul_round_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule 0 (lower (has_type (ty_vec_fits_in_register ty) (sqmul_round_sat x y)))
+  (rv_vsmul_vv x y (unmasked) ty))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (sqmul_round_sat x (splat y))))
+  (rv_vsmul_vx x y (unmasked) ty))
+
+(rule 2 (lower (has_type (ty_vec_fits_in_register ty) (sqmul_round_sat (splat x) y)))
+  (rv_vsmul_vx y x (unmasked) ty))
+
+;;;; Rules for `snarrow` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type (ty_vec_fits_in_register out_ty) (snarrow x @ (value_type in_ty) y)))
+  (if-let lane_diff (u64_to_uimm5 (u64_udiv (ty_lane_count out_ty) 2)))
+  (if-let zero (u64_to_uimm5 0))
+  (let ((x_clip VReg (rv_vnclip_wi x zero (unmasked) (vstate_mf2 (ty_half_lanes out_ty))))
+        (y_clip VReg (rv_vnclip_wi y zero (unmasked) (vstate_mf2 (ty_half_lanes out_ty)))))
+    (rv_vslideup_vvi x_clip y_clip lane_diff (unmasked) out_ty)))
+
+;;;; Rules for `uunarrow` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type (ty_vec_fits_in_register out_ty) (uunarrow x @ (value_type in_ty) y)))
+  (if-let lane_diff (u64_to_uimm5 (u64_udiv (ty_lane_count out_ty) 2)))
+  (if-let zero (u64_to_uimm5 0))
+  (let ((x_clip VReg (rv_vnclipu_wi x zero (unmasked) (vstate_mf2 (ty_half_lanes out_ty))))
+        (y_clip VReg (rv_vnclipu_wi y zero (unmasked) (vstate_mf2 (ty_half_lanes out_ty)))))
+    (rv_vslideup_vvi x_clip y_clip lane_diff (unmasked) out_ty)))
+
+;;;; Rules for `unarrow` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; We don't have a instruction that saturates a signed source into an unsigned destination.
+;; To correct for this we just remove negative values using `vmax` and then use the normal
+;; unsigned to unsigned narrowing instruction.
+
+(rule (lower (has_type (ty_vec_fits_in_register out_ty) (unarrow x @ (value_type in_ty) y)))
+  (if-let lane_diff (u64_to_uimm5 (u64_udiv (ty_lane_count out_ty) 2)))
+  (if-let zero (u64_to_uimm5 0))
+  (let ((x_pos VReg (rv_vmax_vx x (zero_reg) (unmasked) in_ty))
+        (y_pos VReg (rv_vmax_vx y (zero_reg) (unmasked) in_ty))
+        (x_clip VReg (rv_vnclipu_wi x_pos zero (unmasked) (vstate_mf2 (ty_half_lanes out_ty))))
+        (y_clip VReg (rv_vnclipu_wi y_pos zero (unmasked) (vstate_mf2 (ty_half_lanes out_ty)))))
+    (rv_vslideup_vvi x_clip y_clip lane_diff (unmasked) out_ty)))