[Cranelift] add simplification rules

myunbin · myunbin · commit b60f059eadae · 2026-07-09T07:27:29.000Z
diff --git a/cranelift/codegen/src/opts/arithmetic.isle b/cranelift/codegen/src/opts/arithmetic.isle
@@ -568,3 +568,12 @@
 
 ;; (-X) * C = X * (-C)
 (rule (simplify (imul (fits_in_64 ty) (ineg ty x) (iconst ty y))) (imul ty x (iconst ty (imm64_neg ty y))))
+
+;; -(~x) --> x + 1
+(rule (simplify (ineg ty (bnot ty x))) (iadd ty x (iconst_u ty 1)))
+
+;; ~(-x) --> x - 1.
+(rule (simplify (bnot ty (ineg ty x))) (isub ty x (iconst_u ty 1)))
+
+;; -1 - x --> ~x
+(rule (simplify (isub ty (iconst_s ty -1) x)) (bnot ty x))
diff --git a/cranelift/codegen/src/opts/bitops.isle b/cranelift/codegen/src/opts/bitops.isle
@@ -15,7 +15,7 @@
       (subsume x))
 
 ;; x ^ x == 0.
-(rule (simplify (bxor (ty_int ty) x x))
+(rule (simplify (bxor ty x x))
       (subsume (iconst_u ty 0)))
 
 ;; x ^ not(x) == not(x) ^ x == x | not(x) == not(x) | x == -1.
diff --git a/cranelift/codegen/src/opts/extends.isle b/cranelift/codegen/src/opts/extends.isle
@@ -93,3 +93,6 @@
 ;; Try to transform an `iconcat` into an i128 into either an sextend or uextend
 (rule (simplify (iconcat $I128 x (iconst_u _ 0))) (uextend $I128 x))
 (rule (simplify (iconcat $I128 x (sshr _ x (iconst_u _ 63)))) (sextend $I128 x))
+
+;; Select narrowest type
+(rule (simplify (ireduce ty (ireduce cty x))) (ireduce ty x))
diff --git a/cranelift/codegen/src/opts/icmp.isle b/cranelift/codegen/src/opts/icmp.isle
@@ -2,16 +2,16 @@
 
 ;; `x == x` is always true for integers; `x != x` is false. Strict
 ;; inequalities are false, and loose inequalities are true.
-(rule (simplify (eq  (ty_int ty) x x)) (subsume (iconst_u ty 1)))
-(rule (simplify (ne  (ty_int ty) x x)) (subsume (iconst_u ty 0)))
-(rule (simplify (ugt (ty_int ty) x x)) (subsume (iconst_u ty 0)))
-(rule (simplify (uge (ty_int ty) x x)) (subsume (iconst_u ty 1)))
-(rule (simplify (sgt (ty_int ty) x x)) (subsume (iconst_u ty 0)))
-(rule (simplify (sge (ty_int ty) x x)) (subsume (iconst_u ty 1)))
-(rule (simplify (ult (ty_int ty) x x)) (subsume (iconst_u ty 0)))
-(rule (simplify (ule (ty_int ty) x x)) (subsume (iconst_u ty 1)))
-(rule (simplify (slt (ty_int ty) x x)) (subsume (iconst_u ty 0)))
-(rule (simplify (sle (ty_int ty) x x)) (subsume (iconst_u ty 1)))
+(rule (simplify (eq ty x x)) (subsume (cmp_true ty)))
+(rule (simplify (ne ty x x)) (subsume (iconst_u ty 0)))
+(rule (simplify (ugt ty x x)) (subsume (iconst_u ty 0)))
+(rule (simplify (uge ty x x)) (subsume (cmp_true ty)))
+(rule (simplify (sgt ty x x)) (subsume (iconst_u ty 0)))
+(rule (simplify (sge ty x x)) (subsume (cmp_true ty)))
+(rule (simplify (ult ty x x)) (subsume (iconst_u ty 0)))
+(rule (simplify (ule ty x x)) (subsume (cmp_true ty)))
+(rule (simplify (slt ty x x)) (subsume (iconst_u ty 0)))
+(rule (simplify (sle ty x x)) (subsume (cmp_true ty)))
 
 ;; For integers, adding the same thing on both sides of an equality check
 ;; (or an inequality check) doesn't change the result.
@@ -474,3 +474,9 @@
 
 (rule (simplify_skeleton (brif (ireduce _ (clz x_ty x)) _ _))
       (replace_branch_cond (sge $I8 x (iconst_u x_ty 0))))
+
+;; ~x == x --> 0
+(rule (simplify (eq ty (bnot cty x) x)) (subsume (iconst_u ty 0)))
+
+;; ~x != x --> 1
+(rule (simplify (ne ty (bnot cty x) x)) (subsume (cmp_true ty)))
diff --git a/cranelift/codegen/src/opts/shifts.isle b/cranelift/codegen/src/opts/shifts.isle
@@ -315,3 +315,15 @@
 (rule (simplify (iadd ty (ishl ty x z) (ishl ty y z))) (ishl ty (iadd ty x y) z))
 
 (rule (simplify (ushr ty (band ty (ishl ty x y) z) y)) (band ty x (ushr ty z y)))
+
+;; Zero remains zero for any shift or rotate amount.
+(rule (simplify (ishl ty (iconst_u ty 0) x)) (subsume (iconst_u ty 0)))
+(rule (simplify (ushr ty (iconst_u ty 0) x)) (subsume (iconst_u ty 0)))
+(rule (simplify (sshr ty (iconst_u ty 0) x)) (subsume (iconst_u ty 0)))
+(rule (simplify (rotl ty (iconst_u ty 0) x)) (subsume (iconst_u ty 0)))
+(rule (simplify (rotr ty (iconst_u ty 0) x)) (subsume (iconst_u ty 0)))
+
+;; All ones remain all ones for arithmetic right shifts or rotates.
+(rule (simplify (sshr ty (iconst_s ty -1) x)) (subsume (iconst_s ty -1)))
+(rule (simplify (rotl ty (iconst_s ty -1) x)) (subsume (iconst_s ty -1)))
+(rule (simplify (rotr ty (iconst_s ty -1) x)) (subsume (iconst_s ty -1)))
diff --git a/cranelift/filetests/filetests/egraph/arithmetic-precise.clif b/cranelift/filetests/filetests/egraph/arithmetic-precise.clif
@@ -247,8 +247,8 @@ block0(v0: i32x4, v1: i32x4):
 ;     const0 = 0xffffffffffffffffffffffffffffffff
 ;
 ; block0(v0: i32x4, v1: i32x4):
-;     v9 = icmp eq v0, v0
-;     return v9
+;     v7 = vconst.i32x4 const0
+;     return v7  ; v7 = const0
 ; }
 
 ;; (x - y) != x --> y != 0
diff --git a/cranelift/filetests/filetests/egraph/arithmetic.clif b/cranelift/filetests/filetests/egraph/arithmetic.clif
@@ -219,6 +219,17 @@ block0(v0: i32):
 ; check: v6 = ineg v0
 ; check: return v6
 
+function %ineg_bnot_to_iadd_one(i32) -> i32 {
+block0(v0: i32):
+    v1 = bnot v0
+    v2 = ineg v1
+    return v2
+}
+
+; check: v3 = iconst.i32 1
+; check: v4 = iadd v0, v3  ; v3 = 1
+; check: return v4
+
 function %byte_sub_smax_twice(i8) -> i8 {
 block0(v0: i8):
     v1 = iconst.i8 127
@@ -464,3 +475,13 @@ block0(v0: i64):
     ; check: v6 = iadd v4, v5
     ; check: return v6
 }
+
+;; -1 - x --> ~x
+function %isub_minus_one_lhs_to_bnot(i32) -> i32 {
+block0(v0: i32):
+    v1 = iconst.i32 -1
+    v2 = isub v1, v0
+    return v2
+    ; check: v12 = bnot v0
+    ; check: return v12
+}
diff --git a/cranelift/filetests/filetests/egraph/bitops.clif b/cranelift/filetests/filetests/egraph/bitops.clif
@@ -160,6 +160,17 @@ block0(v0: i32x4):
 ; check: v5 = vconst.i32x4 const0
 ; check: return v5
 
+function %vector_bxor_bnot_x_bnot_x(i32x4) -> i32x4 {
+block0(v0: i32x4):
+    v1 = bnot v0
+    v2 = bxor v1, v1
+    return v2
+}
+
+; check: const0 = 0x00000000000000000000000000000000
+; check: v5 = vconst.i32x4 const0
+; check: return v5
+
 function %vector_bor_x_bnot_x(i32x4) -> i32x4 {
 block0(v0: i32x4):
     v1 = bnot v0
@@ -803,6 +814,21 @@ block0(v0: i32, v1: i32):
 ;     return v5
 ; }
 
+;; (bnot ty (ineg ty x)) -> (isub ty x (iconst_u ty 1))
+function %test_bnot_ineg_to_isub_one(i32) -> i32 fast {
+block0(v0: i32):
+    v1 = ineg v0
+    v2 = bnot v1
+    return v2
+}
+
+; function %test_bnot_ineg_to_isub_one(i32) -> i32 fast {
+; block0(v0: i32):
+;     v3 = iconst.i32 1
+;     v4 = isub v0, v3
+;     return v4
+; }
+
 ;; (bor ty (bxor ty x z) (bor ty y x)) -> (bor ty (bor ty y x) z)
 function %test_bor_bxor_bor(i32, i32, i32) -> i32 fast {
 block0(v0: i32, v1: i32, v2: i32):
diff --git a/cranelift/filetests/filetests/egraph/extends.clif b/cranelift/filetests/filetests/egraph/extends.clif
@@ -134,6 +134,16 @@ block0(v1: i16):
 ; check: v4 = uextend.i32 v1
 ; check: return v4
 
+function %nested_ireduce(i64) -> i8 {
+block0(v1: i64):
+    v2 = ireduce.i32 v1
+    v3 = ireduce.i8 v2
+    return v3
+}
+
+; check: v4 = ireduce.i8 v1
+; check: return v4
+
 function %sextend_then_slt_zero(i8) -> i8 {
 block0(v0: i8):
     v1 = sextend.i16 v0
diff --git a/cranelift/filetests/filetests/egraph/i128-opts.clif b/cranelift/filetests/filetests/egraph/i128-opts.clif
@@ -74,16 +74,18 @@ function %eq_self_v128(i8x16) -> i8x16 {
 block0(v0: i8x16):
     v1 = icmp eq v0, v0
     return v1
-    ; check: v1 = icmp eq v0, v0
-    ; check: return v1
+    ; check: const0 = 0xffffffffffffffffffffffffffffffff
+    ; check: v4 = vconst.i8x16 const0
+    ; check: return v4  ; v4 = const0
 }
 
 function %ne_self_v128(i8x16) -> i8x16 {
 block0(v0: i8x16):
     v1 = icmp ne v0, v0
     return v1
-    ; check: v1 = icmp ne v0, v0
-    ; check: return v1
+    ; check: const0 = 0x00000000000000000000000000000000
+    ; check: v4 = vconst.i8x16 const0
+    ; check: return v4  ; v4 = const0
 }
 
 function %bor_extended_constants_i128() -> i128 {
diff --git a/cranelift/filetests/filetests/egraph/icmp.clif b/cranelift/filetests/filetests/egraph/icmp.clif
@@ -202,6 +202,34 @@ block0(v0: i32, v1: i32):
 ;     return v5
 ; }
 
+;; `~x == x` to false.
+function %simplify_icmp_eq_bnot_x_x(i32) -> i8 fast {
+block0(v0: i32):
+    v1 = bnot v0
+    v2 = icmp eq v1, v0
+    return v2
+}
+
+; function %simplify_icmp_eq_bnot_x_x(i32) -> i8 fast {
+; block0(v0: i32):
+;     v3 = iconst.i8 0
+;     return v3  ; v3 = 0
+; }
+
+;; `~x != x` to true.
+function %simplify_icmp_ne_bnot_x_x(i32) -> i8 fast {
+block0(v0: i32):
+    v1 = bnot v0
+    v2 = icmp ne v1, v0
+    return v2
+}
+
+; function %simplify_icmp_ne_bnot_x_x(i32) -> i8 fast {
+; block0(v0: i32):
+;     v3 = iconst.i8 1
+;     return v3  ; v3 = 1
+; }
+
 function %issue_10929_no_crash_on_icmp_vectors() -> i32x4 {
   const0 = 0x40ad3fb47cb16076c8cb1fdd8189d40f
 
diff --git a/cranelift/filetests/filetests/egraph/shifts.clif b/cranelift/filetests/filetests/egraph/shifts.clif
@@ -606,3 +606,66 @@ block0(v0: i64):
     v2 = ishl v0, v1
     return v2
 }
+function %ishl_zero_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 0
+    v2 = ishl v1, v0
+    return v2
+    ; check: return v1
+}
+function %ushr_zero_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 0
+    v2 = ushr v1, v0
+    return v2
+    ; check: return v1
+}
+function %rotl_zero_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 0
+    v2 = rotl v1, v0
+    return v2
+    ; check: return v1
+}
+function %rotr_zero_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 0
+    v2 = rotr v1, v0
+    return v2
+    ; check: return v1
+}
+function %rotl_const_multiple_of_64(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 128
+    v2 = rotl v0, v1
+    return v2
+    ; check: return v0
+}
+function %sshr_all_ones_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 -1
+    v2 = sshr v1, v0
+    return v2
+    ; check: return v1
+}
+function %rotl_all_ones_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 -1
+    v2 = rotl v1, v0
+    return v2
+    ; check: return v1
+}
+function %rotr_all_ones_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 -1
+    v2 = rotr v1, v0
+    return v2
+    ; check: return v1
+}
+function %sshr_zero_left_operand(i64) -> i64 {
+block0(v0: i64):
+    v1 = iconst.i64 0
+    v2 = sshr v1, v0
+    return v2
+    ; check: return v1
+}
diff --git a/cranelift/filetests/filetests/runtests/arithmetic.clif b/cranelift/filetests/filetests/runtests/arithmetic.clif
@@ -1065,3 +1065,15 @@ block0(v0: i64):
 ; run: %fold_bnot_iadd_neg_const_i64_rt(7) == -1
 ; run: %fold_bnot_iadd_neg_const_i64_rt(-1) == 7
 ; run: %fold_bnot_iadd_neg_const_i64_rt(0x80000000_00000000) == 0x80000000_00000006
+
+function %fold_isub_minus_one_lhs_to_bnot_i32_rt(i32) -> i32 fast {
+block0(v0: i32):
+    v1 = iconst.i32 -1
+    v2 = isub v1, v0
+    return v2
+}
+
+; run: %fold_isub_minus_one_lhs_to_bnot_i32_rt(0) == -1
+; run: %fold_isub_minus_one_lhs_to_bnot_i32_rt(1) == -2
+; run: %fold_isub_minus_one_lhs_to_bnot_i32_rt(-1) == 0
+; run: %fold_isub_minus_one_lhs_to_bnot_i32_rt(0x7fffffff) == 0x80000000
diff --git a/cranelift/filetests/filetests/runtests/bitops.clif b/cranelift/filetests/filetests/runtests/bitops.clif
@@ -554,6 +554,17 @@ block0(v0: i32, v1: i32):
 ; run: %test_bnot_iadd_bnot_to_isub_rt(4, 5) == -1
 ; run: %test_bnot_iadd_bnot_to_isub_rt(0x12345678, 0x11111111) == 0x01234567
 
+function %test_bnot_ineg_to_isub_one_rt(i32) -> i32 fast {
+block0(v0: i32):
+    v1 = ineg v0
+    v2 = bnot v1
+    return v2
+}
+
+; run: %test_bnot_ineg_to_isub_one_rt(0) == -1
+; run: %test_bnot_ineg_to_isub_one_rt(1) == 0
+; run: %test_bnot_ineg_to_isub_one_rt(0x12345678) == 0x12345677
+
 function %test_bor_bxor_bor(i32, i32, i32) -> i32 fast {
 block0(v0: i32, v1: i32, v2: i32):
     v3 = bxor v0, v2
@@ -807,4 +818,4 @@ block0(v0: i32, v1: i32):
 }
 
 ; run: %test_ule_x_bor(0, 0) == 1
-; run: %test_ule_x_bor(1, 0) == 1
+; run: %test_ule_x_bor(1, 0) == 1
diff --git a/cranelift/filetests/filetests/runtests/extend.clif b/cranelift/filetests/filetests/runtests/extend.clif
@@ -238,3 +238,13 @@ block0(v0: i8):
 }
 ; run: %zext16_sext32(0xff) == 0xff
 ; run: %zext16_sext32(0x7f) == 0x7f
+
+function %nested_ireduce_i64_i8(i64) -> i8 {
+block0(v0: i64):
+    v1 = ireduce.i32 v0
+    v2 = ireduce.i8 v1
+    return v2
+}
+; run: %nested_ireduce_i64_i8(0) == 0
+; run: %nested_ireduce_i64_i8(-1) == -1
+; run: %nested_ireduce_i64_i8(0x1234_5678_9abc_def0) == 0xf0
diff --git a/cranelift/filetests/filetests/runtests/icmp.clif b/cranelift/filetests/filetests/runtests/icmp.clif
@@ -65,6 +65,28 @@ block0(v0: i8, v2: i8):
 
 ; run: %fold_cmp_over_nots2(2, 1) == 1
 
+function %eq_bnot_x_x(i32) -> i8 fast {
+block0(v0: i32):
+    v1 = bnot v0
+    v2 = icmp eq v1, v0
+    return v2
+}
+
+; run: %eq_bnot_x_x(0) == 0
+; run: %eq_bnot_x_x(-1) == 0
+; run: %eq_bnot_x_x(42) == 0
+
+function %ne_bnot_x_x(i32) -> i8 fast {
+block0(v0: i32):
+    v1 = bnot v0
+    v2 = icmp ne v1, v0
+    return v2
+}
+
+; run: %ne_bnot_x_x(0) == 1
+; run: %ne_bnot_x_x(-1) == 1
+; run: %ne_bnot_x_x(42) == 1
+
 ;; (bxor ty (slt cty x y) (sgt cty x y)) -> (ne cty x y)
 function %test_bxor_slt_sgt(i32, i32) -> i8 fast {
 block0(v0: i32, v1: i32):
diff --git a/cranelift/filetests/filetests/runtests/ineg.clif b/cranelift/filetests/filetests/runtests/ineg.clif
diff --git a/cranelift/filetests/filetests/runtests/rotl.clif b/cranelift/filetests/filetests/runtests/rotl.clif
diff --git a/cranelift/filetests/filetests/runtests/shifts.clif b/cranelift/filetests/filetests/runtests/shifts.clif
diff --git a/cranelift/filetests/filetests/runtests/simd-vector.clif b/cranelift/filetests/filetests/runtests/simd-vector.clif
diff --git a/tests/disas/gc/copying/v128-fields.wat b/tests/disas/gc/copying/v128-fields.wat
diff --git a/tests/disas/gc/drc/externref-globals.wat b/tests/disas/gc/drc/externref-globals.wat
diff --git a/tests/disas/gc/drc/struct-set.wat b/tests/disas/gc/drc/struct-set.wat
diff --git a/tests/disas/gc/null/v128-fields.wat b/tests/disas/gc/null/v128-fields.wat
