Cranelift: add mid-end opt rules

cranelift/codegen/src/egraph/cost.rs

@@ -79,6 +79,7 @@ impl Cost {
             // "Simple" arithmetic.
             Opcode::Iadd
             | Opcode::Isub
+            | Opcode::Ineg
             | Opcode::Band
             | Opcode::Bor
             | Opcode::Bxor

cranelift/codegen/src/opts/arithmetic.isle

@@ -699,3 +699,35 @@
                       (isub ty (iconst ty x) y)
                       (iadd ty z (iconst ty p))))
       (iadd ty (isub ty z y) (iconst ty (imm64_add ty p x))))
+
+;; -(x & y) + x --> x & ~y
+(rule (simplify (iadd ty (ineg ty (band ty x y)) x)) (band ty x (bnot ty y)))
+(rule (simplify (iadd ty (ineg ty (band ty x y)) y)) (band ty y (bnot ty x)))
+(rule (simplify (iadd ty x (ineg ty (band ty x y)))) (band ty x (bnot ty y)))
+(rule (simplify (iadd ty x (ineg ty (band ty y x)))) (band ty x (bnot ty y)))
+
+;; (~x & y) + (~y & x) --> y ^ x
+(rule (simplify (iadd ty (band ty (bnot ty x) y) (band ty (bnot ty y) x))) (bxor ty y x))
+(rule (simplify (iadd ty (band ty (bnot ty x) y) (band ty x (bnot ty y)))) (bxor ty y x))
+(rule (simplify (iadd ty (band ty x (bnot ty y)) (band ty (bnot ty x) y))) (bxor ty x y))
+(rule (simplify (iadd ty (band ty x (bnot ty y)) (band ty y (bnot ty x)))) (bxor ty x y))
+
+;; max(~x, ~y) --> ~min(x, y)
+(rule (simplify (smax ty (bnot ty x) (bnot ty y))) (bnot ty (smin ty x y)))
+(rule (simplify (umax ty (bnot ty x) (bnot ty y))) (bnot ty (umin ty x y)))
+
+;; min(~x, ~y) --> ~max(x, y)
+(rule (simplify (smin ty (bnot ty x) (bnot ty y))) (bnot ty (smax ty x y)))
+(rule (simplify (umin ty (bnot ty x) (bnot ty y))) (bnot ty (umax ty x y)))
+
+;; (-x) - (~x) --> 1
+(rule (simplify (isub (ty_int_vec128 ty) (ineg ty x) (bnot ty x))) (iconst_u ty 1))
+
+;; ~x + C --> (C - 1) - x
+(rule (simplify (iadd (fits_in_64 ty) (bnot ty x) (iconst ty y))) (isub ty (iconst ty (imm64_sub ty y (imm64 1))) x))
+
+;; (x | C) - C --> x & ~C
+(rule (simplify (isub (fits_in_64 ty) (bor ty x (iconst ty y)) (iconst ty y))) (band ty x (iconst ty (imm64_not ty y))))
+
+;; (x << 1) * C --> x * (2 * C)
+(rule (simplify (imul (fits_in_64 ty) (ishl ty x (iconst_u ty 1)) (iconst ty y))) (imul ty x (iconst ty (imm64_mul ty y (imm64 2)))))

cranelift/codegen/src/opts/bitops.isle

@@ -846,3 +846,82 @@
 
 ;; ~(x & y) | x --> -1
 (rule (simplify (bor (ty_int_vec128 ty) (bnot ty (band ty x y)) x)) (subsume (iconst_s ty -1)))
+
+;; (~x | y) ^ x --> ~(x & y)
+(rule (simplify (bxor ty (bor ty (bnot ty x) y) x)) (bnot ty (band ty x y)))
+(rule (simplify (bxor ty (bor ty x (bnot ty y)) y)) (bnot ty (band ty y x)))
+(rule (simplify (bxor ty x (bor ty (bnot ty x) y))) (bnot ty (band ty x y)))
+(rule (simplify (bxor ty x (bor ty y (bnot ty x)))) (bnot ty (band ty x y)))
+
+;; max(x, y) & min(x, y) --> x & y
+(rule (simplify (band ty (smax ty x y) (smin ty x y))) (band ty x y))
+(rule (simplify (band ty (smax ty x y) (smin ty y x))) (band ty x y))
+(rule (simplify (band ty (smin ty x y) (smax ty x y))) (band ty x y))
+(rule (simplify (band ty (smin ty x y) (smax ty y x))) (band ty x y))
+(rule (simplify (band ty (umax ty x y) (umin ty x y))) (band ty x y))
+(rule (simplify (band ty (umax ty x y) (umin ty y x))) (band ty x y))
+(rule (simplify (band ty (umin ty x y) (umax ty x y))) (band ty x y))
+(rule (simplify (band ty (umin ty x y) (umax ty y x))) (band ty x y))
+
+;; max(x, y) | min(x, y) --> x | y
+(rule (simplify (bor ty (smax ty x y) (smin ty x y))) (bor ty x y))
+(rule (simplify (bor ty (smax ty x y) (smin ty y x))) (bor ty x y))
+(rule (simplify (bor ty (smin ty x y) (smax ty x y))) (bor ty x y))
+(rule (simplify (bor ty (smin ty x y) (smax ty y x))) (bor ty x y))
+(rule (simplify (bor ty (umax ty x y) (umin ty x y))) (bor ty x y))
+(rule (simplify (bor ty (umax ty x y) (umin ty y x))) (bor ty x y))
+(rule (simplify (bor ty (umin ty x y) (umax ty x y))) (bor ty x y))
+(rule (simplify (bor ty (umin ty x y) (umax ty y x))) (bor ty x y))
+
+;; max(x, y) ^ min(x, y) --> x ^ y
+(rule (simplify (bxor ty (smax ty x y) (smin ty x y))) (bxor ty x y))
+(rule (simplify (bxor ty (smax ty x y) (smin ty y x))) (bxor ty x y))
+(rule (simplify (bxor ty (smin ty x y) (smax ty x y))) (bxor ty x y))
+(rule (simplify (bxor ty (smin ty x y) (smax ty y x))) (bxor ty x y))
+(rule (simplify (bxor ty (umax ty x y) (umin ty x y))) (bxor ty x y))
+(rule (simplify (bxor ty (umax ty x y) (umin ty y x))) (bxor ty x y))
+(rule (simplify (bxor ty (umin ty x y) (umax ty x y))) (bxor ty x y))
+(rule (simplify (bxor ty (umin ty x y) (umax ty y x))) (bxor ty x y))
+
+;; (~x OP y) ^ (~y OP x) --> x ^ y (OP = {&, |})
+(rule (simplify (bxor ty (bor ty (bnot ty x) y) (bor ty (bnot ty y) x))) (bxor ty x y))
+(rule (simplify (bxor ty (bor ty (bnot ty x) y) (bor ty x (bnot ty y)))) (bxor ty x y))
+(rule (simplify (bxor ty (bor ty x (bnot ty y)) (bor ty (bnot ty x) y))) (bxor ty y x))
+(rule (simplify (bxor ty (bor ty x (bnot ty y)) (bor ty y (bnot ty x)))) (bxor ty y x))
+(rule (simplify (bxor ty (band ty (bnot ty x) y) (band ty (bnot ty y) x))) (bxor ty y x))
+(rule (simplify (bxor ty (band ty (bnot ty x) y) (band ty x (bnot ty y)))) (bxor ty y x))
+(rule (simplify (bxor ty (band ty x (bnot ty y)) (band ty (bnot ty x) y))) (bxor ty x y))
+(rule (simplify (bxor ty (band ty x (bnot ty y)) (band ty y (bnot ty x)))) (bxor ty x y))
+
+;; (~x | y) & (~y | x) --> ~(x ^ y)
+(rule (simplify (band ty (bor ty (bnot ty x) y) (bor ty (bnot ty y) x))) (bnot ty (bxor ty x y)))
+(rule (simplify (band ty (bor ty (bnot ty x) y) (bor ty x (bnot ty y)))) (bnot ty (bxor ty x y)))
+(rule (simplify (band ty (bor ty x (bnot ty y)) (bor ty (bnot ty x) y))) (bnot ty (bxor ty y x)))
+(rule (simplify (band ty (bor ty x (bnot ty y)) (bor ty y (bnot ty x)))) (bnot ty (bxor ty y x)))
+
+;; (x + C) ^ -1 --> (-(C + 1)) - x
+(rule (simplify (bxor (fits_in_64 ty) (iadd ty x (iconst ty y)) (iconst_s ty -1))) (isub ty (ineg ty (iconst ty (imm64_add ty y (imm64 1)))) x))
+
+;; (x | C) ^ (y | C) --> (x ^ y) & ~C
+(rule (simplify (bxor (fits_in_64 (ty_int ty)) (bor ty (iconst ty x) y) (bor ty z (iconst ty x)))) (band ty (bxor ty z y) (iconst ty (imm64_not ty x))))
+(rule (simplify (bxor (fits_in_64 (ty_int ty)) (bor ty x (iconst ty y)) (bor ty z (iconst ty y)))) (band ty (bxor ty x z) (iconst ty (imm64_not ty y))))
+
+;; (x | C) ^ (y ^ C) --> (x & ~C) ^ y
+(rule (simplify (bxor (fits_in_64 (ty_int ty)) (bor ty x (iconst ty y)) (bxor ty z (iconst ty y)))) (bxor ty (band ty x (iconst ty (imm64_not ty y))) z))
+(rule (simplify (bxor (fits_in_64 (ty_int ty)) (bxor ty x (iconst ty y)) (bor ty z (iconst ty y)))) (bxor ty (band ty z (iconst ty (imm64_not ty y))) x))
+
+;; ~smax(~x, y) --> smin(x, ~y)
+(rule (simplify (bnot ty (smax ty (bnot ty x) y))) (smin ty x (bnot ty y)))
+(rule (simplify (bnot ty (smax ty x (bnot ty y)))) (smin ty y (bnot ty x)))
+
+;; ~umax(~x, y) --> umin(x, ~y)
+(rule (simplify (bnot ty (umax ty (bnot ty x) y))) (umin ty x (bnot ty y)))
+(rule (simplify (bnot ty (umax ty x (bnot ty y)))) (umin ty y (bnot ty x)))
+
+;; ~smin(~x, y) --> smax(x, ~y)
+(rule (simplify (bnot ty (smin ty (bnot ty x) y))) (smax ty x (bnot ty y)))
+(rule (simplify (bnot ty (smin ty x (bnot ty y)))) (smax ty y (bnot ty x)))
+
+;; ~umin(~x, y) --> umax(x, ~y)
+(rule (simplify (bnot ty (umin ty (bnot ty x) y))) (umax ty x (bnot ty y)))
+(rule (simplify (bnot ty (umin ty x (bnot ty y)))) (umax ty y (bnot ty x)))

cranelift/codegen/src/opts/icmp.isle

@@ -628,3 +628,52 @@
 (rule (simplify (bor _ (ne cty x (iconst_u cty 0)) (ne cty y (iconst_u cty 0)))) (ne cty (bor cty x y) (all_zero cty)))
 ;; (x == 0) & (y == 0) --> (x | y) == 0
 (rule (simplify (band _ (eq cty x (iconst_u cty 0)) (eq cty y (iconst_u cty 0)))) (eq cty (bor cty x y) (all_zero cty)))
+
+;; (x << 1) == x --> x == 0
+(rule (simplify (eq ty (ishl cty x (iconst_u _ 1)) x)) (eq ty x (all_zero cty)))
+(rule (simplify (eq ty x (ishl cty x (iconst_u _ 1)))) (eq ty x (all_zero cty)))
+
+;; (x >>u 1) == x --> x == 0
+(rule (simplify (eq ty (ushr cty x (iconst_u _ 1)) x)) (eq ty x (all_zero cty)))
+(rule (simplify (eq ty x (ushr cty x (iconst_u _ 1)))) (eq ty x (all_zero cty)))
+
+;; (~x & y) == y --> (y & x) == 0
+(rule (simplify (eq ty (band cty (bnot cty x) y) y)) (eq ty (band cty y x) (all_zero cty)))
+(rule (simplify (eq ty (band cty x (bnot cty y)) x)) (eq ty (band cty x y) (all_zero cty)))
+(rule (simplify (eq ty x (band cty (bnot cty y) x))) (eq ty (band cty x y) (all_zero cty)))
+(rule (simplify (eq ty x (band cty x (bnot cty y)))) (eq ty (band cty x y) (all_zero cty)))
+
+;; umax(x, y) OP 0 --> (x | y) OP 0 (OP = {==, !=})
+(rule (simplify (eq ty (umax cty x y) (all_zero cty))) (eq ty (bor cty x y) (all_zero cty)))
+(rule (simplify (ne ty (umax cty x y) (all_zero cty))) (ne ty (bor cty x y) (all_zero cty)))
+
+;; umax(x, y) OP umin(x, y) --> x OP y (OP = {==, !=})
+(rule (simplify (eq ty (umax cty x y) (umin cty x y))) (eq ty x y))
+(rule (simplify (eq ty (umax cty x y) (umin cty y x))) (eq ty x y))
+(rule (simplify (eq ty (umin cty x y) (umax cty x y))) (eq ty x y))
+(rule (simplify (eq ty (umin cty x y) (umax cty y x))) (eq ty x y))
+(rule (simplify (ne ty (umax cty x y) (umin cty x y))) (ne ty x y))
+(rule (simplify (ne ty (umax cty x y) (umin cty y x))) (ne ty x y))
+(rule (simplify (ne ty (umin cty x y) (umax cty x y))) (ne ty x y))
+(rule (simplify (ne ty (umin cty x y) (umax cty y x))) (ne ty x y))
+
+;; ~x == (~x | y) --> (y & x) == 0
+(rule (simplify (eq ty (bnot cty x) (bor cty (bnot cty x) y))) (eq ty (band cty y x) (all_zero cty)))
+(rule (simplify (eq ty (bnot cty x) (bor cty y (bnot cty x)))) (eq ty (band cty y x) (all_zero cty)))
+(rule (simplify (eq ty (bor cty (bnot cty x) y) (bnot cty x))) (eq ty (band cty y x) (all_zero cty)))
+(rule (simplify (eq ty (bor cty x (bnot cty y)) (bnot cty y))) (eq ty (band cty x y) (all_zero cty)))
+
+;; (c ? x : y) == (c ? y : x) --> x == y
+(rule (simplify (eq ty (select cty x y z) (select cty x z y))) (eq ty y z))
+
+;; (x >=s y) & (y >=s x) --> x == y
+(rule (simplify (band _ (sge cty x y) (sge cty y x))) (eq cty x y))
+(rule (simplify (band _ (sge cty x y) (sle cty x y))) (eq cty x y))
+(rule (simplify (band _ (sle cty x y) (sge cty x y))) (eq cty x y))
+(rule (simplify (band _ (sle cty x y) (sle cty y x))) (eq cty x y))
+
+;; (x >=u y) & (y >=u x) --> x == y
+(rule (simplify (band _ (uge cty x y) (uge cty y x))) (eq cty x y))
+(rule (simplify (band _ (uge cty x y) (ule cty x y))) (eq cty x y))
+(rule (simplify (band _ (ule cty x y) (uge cty x y))) (eq cty x y))
+(rule (simplify (band _ (ule cty x y) (ule cty y x))) (eq cty x y))

cranelift/codegen/src/opts/selects.isle

@@ -313,3 +313,25 @@
 
 ;; (c ? C1 : C2) - C --> c ? (C1 - C) : (C2 - C)
 (rule (simplify (isub (fits_in_64 ty) (select ty x (iconst ty y) (iconst ty z)) (iconst ty p))) (select ty x (iconst ty (imm64_sub ty y p)) (iconst ty (imm64_sub ty z p))))
+
+;; (x == 0) ? -x : x --> x
+(rule (simplify (select ty (eq cty x (all_zero ty)) (ineg ty x) x)) x)
+(rule (simplify (select ty (ne cty x (all_zero ty)) x (ineg ty x))) x)
+
+;; ~(c ? C1 : C2) --> c ? ~C1 : ~C2
+(rule (simplify (bnot (fits_in_64 ty) (select ty x (iconst ty y) (iconst ty z)))) (select ty x (iconst ty (imm64_not ty y)) (iconst ty (imm64_not ty z))))
+
+;; -(c ? C1 : C2) --> c ? -C1 : -C2
+(rule (simplify (ineg (fits_in_64 ty) (select ty x (iconst ty y) (iconst ty z)))) (select ty x (iconst ty (imm64_neg ty y)) (iconst ty (imm64_neg ty z))))
+
+;; (x == 0) ? 0 : (x & y) --> x & y
+(rule (simplify (select ty (eq cty x (iconst_u ty 0)) (iconst_u ty 0) (band ty x y))) (band ty x y))
+(rule (simplify (select ty (eq cty x (iconst_u ty 0)) (iconst_u ty 0) (band ty y x))) (band ty x y))
+(rule (simplify (select ty (ne cty x (iconst_u ty 0)) (band ty x y) (iconst_u ty 0))) (band ty x y))
+(rule (simplify (select ty (ne cty x (iconst_u ty 0)) (band ty y x) (iconst_u ty 0))) (band ty x y))
+
+;; (x == 0) ? 0 : (x * y) --> x * y
+(rule (simplify (select ty (eq cty x (iconst_u ty 0)) (iconst_u ty 0) (imul ty x y))) (imul ty x y))
+(rule (simplify (select ty (eq cty x (iconst_u ty 0)) (iconst_u ty 0) (imul ty y x))) (imul ty x y))
+(rule (simplify (select ty (ne cty x (iconst_u ty 0)) (imul ty x y) (iconst_u ty 0))) (imul ty x y))
+(rule (simplify (select ty (ne cty x (iconst_u ty 0)) (imul ty y x) (iconst_u ty 0))) (imul ty x y))

cranelift/codegen/src/opts/shifts.isle

@@ -368,3 +368,9 @@
 
 ;; ((x & M) << k) & M --> (x << k) & M (M = 2^n - 1)
 (rule (simplify (band ty (ishl ty (band ty x (iconst_u ty y)) z) (iconst_u ty y))) (if-let true (u64_eq (u64_and y (u64_wrapping_add y 1)) 0)) (band ty (ishl ty x z) (iconst_u ty y)))
+
+;; ~rotl(~x, k) --> rotl(x, k)
+(rule (simplify (bnot ty (rotl ty (bnot ty x) y))) (rotl ty x y))
+
+;; ~rotr(~x, k) --> rotr(x, k)
+(rule (simplify (bnot ty (rotr ty (bnot ty x) y))) (rotr ty x y))